1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* A network driver using virtio. 3 * 4 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation 5 */ 6 //#define DEBUG 7 #include <linux/netdevice.h> 8 #include <linux/etherdevice.h> 9 #include <linux/ethtool.h> 10 #include <linux/module.h> 11 #include <linux/virtio.h> 12 #include <linux/virtio_net.h> 13 #include <linux/bpf.h> 14 #include <linux/bpf_trace.h> 15 #include <linux/scatterlist.h> 16 #include <linux/if_vlan.h> 17 #include <linux/slab.h> 18 #include <linux/cpu.h> 19 #include <linux/average.h> 20 #include <linux/filter.h> 21 #include <linux/kernel.h> 22 #include <net/route.h> 23 #include <net/xdp.h> 24 #include <net/net_failover.h> 25 26 static int napi_weight = NAPI_POLL_WEIGHT; 27 module_param(napi_weight, int, 0444); 28 29 static bool csum = true, gso = true, napi_tx = true; 30 module_param(csum, bool, 0444); 31 module_param(gso, bool, 0444); 32 module_param(napi_tx, bool, 0644); 33 34 /* FIXME: MTU in config. */ 35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN) 36 #define GOOD_COPY_LEN 128 37 38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD) 39 40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */ 41 #define VIRTIO_XDP_HEADROOM 256 42 43 /* Separating two types of XDP xmit */ 44 #define VIRTIO_XDP_TX BIT(0) 45 #define VIRTIO_XDP_REDIR BIT(1) 46 47 #define VIRTIO_XDP_FLAG BIT(0) 48 49 /* RX packet size EWMA. The average packet size is used to determine the packet 50 * buffer size when refilling RX rings. As the entire RX ring may be refilled 51 * at once, the weight is chosen so that the EWMA will be insensitive to short- 52 * term, transient changes in packet size. 53 */ 54 DECLARE_EWMA(pkt_len, 0, 64) 55 56 #define VIRTNET_DRIVER_VERSION "1.0.0" 57 58 static const unsigned long guest_offloads[] = { 59 VIRTIO_NET_F_GUEST_TSO4, 60 VIRTIO_NET_F_GUEST_TSO6, 61 VIRTIO_NET_F_GUEST_ECN, 62 VIRTIO_NET_F_GUEST_UFO, 63 VIRTIO_NET_F_GUEST_CSUM, 64 VIRTIO_NET_F_GUEST_USO4, 65 VIRTIO_NET_F_GUEST_USO6, 66 VIRTIO_NET_F_GUEST_HDRLEN 67 }; 68 69 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \ 70 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \ 71 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \ 72 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \ 73 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \ 74 (1ULL << VIRTIO_NET_F_GUEST_USO6)) 75 76 struct virtnet_stat_desc { 77 char desc[ETH_GSTRING_LEN]; 78 size_t offset; 79 }; 80 81 struct virtnet_sq_stats { 82 struct u64_stats_sync syncp; 83 u64 packets; 84 u64 bytes; 85 u64 xdp_tx; 86 u64 xdp_tx_drops; 87 u64 kicks; 88 u64 tx_timeouts; 89 }; 90 91 struct virtnet_rq_stats { 92 struct u64_stats_sync syncp; 93 u64 packets; 94 u64 bytes; 95 u64 drops; 96 u64 xdp_packets; 97 u64 xdp_tx; 98 u64 xdp_redirects; 99 u64 xdp_drops; 100 u64 kicks; 101 }; 102 103 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m) 104 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m) 105 106 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = { 107 { "packets", VIRTNET_SQ_STAT(packets) }, 108 { "bytes", VIRTNET_SQ_STAT(bytes) }, 109 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) }, 110 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) }, 111 { "kicks", VIRTNET_SQ_STAT(kicks) }, 112 { "tx_timeouts", VIRTNET_SQ_STAT(tx_timeouts) }, 113 }; 114 115 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = { 116 { "packets", VIRTNET_RQ_STAT(packets) }, 117 { "bytes", VIRTNET_RQ_STAT(bytes) }, 118 { "drops", VIRTNET_RQ_STAT(drops) }, 119 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) }, 120 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) }, 121 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) }, 122 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) }, 123 { "kicks", VIRTNET_RQ_STAT(kicks) }, 124 }; 125 126 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc) 127 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc) 128 129 /* Internal representation of a send virtqueue */ 130 struct send_queue { 131 /* Virtqueue associated with this send _queue */ 132 struct virtqueue *vq; 133 134 /* TX: fragments + linear part + virtio header */ 135 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 136 137 /* Name of the send queue: output.$index */ 138 char name[16]; 139 140 struct virtnet_sq_stats stats; 141 142 struct napi_struct napi; 143 144 /* Record whether sq is in reset state. */ 145 bool reset; 146 }; 147 148 /* Internal representation of a receive virtqueue */ 149 struct receive_queue { 150 /* Virtqueue associated with this receive_queue */ 151 struct virtqueue *vq; 152 153 struct napi_struct napi; 154 155 struct bpf_prog __rcu *xdp_prog; 156 157 struct virtnet_rq_stats stats; 158 159 /* Chain pages by the private ptr. */ 160 struct page *pages; 161 162 /* Average packet length for mergeable receive buffers. */ 163 struct ewma_pkt_len mrg_avg_pkt_len; 164 165 /* Page frag for packet buffer allocation. */ 166 struct page_frag alloc_frag; 167 168 /* RX: fragments + linear part + virtio header */ 169 struct scatterlist sg[MAX_SKB_FRAGS + 2]; 170 171 /* Min single buffer size for mergeable buffers case. */ 172 unsigned int min_buf_len; 173 174 /* Name of this receive queue: input.$index */ 175 char name[16]; 176 177 struct xdp_rxq_info xdp_rxq; 178 }; 179 180 /* This structure can contain rss message with maximum settings for indirection table and keysize 181 * Note, that default structure that describes RSS configuration virtio_net_rss_config 182 * contains same info but can't handle table values. 183 * In any case, structure would be passed to virtio hw through sg_buf split by parts 184 * because table sizes may be differ according to the device configuration. 185 */ 186 #define VIRTIO_NET_RSS_MAX_KEY_SIZE 40 187 #define VIRTIO_NET_RSS_MAX_TABLE_LEN 128 188 struct virtio_net_ctrl_rss { 189 u32 hash_types; 190 u16 indirection_table_mask; 191 u16 unclassified_queue; 192 u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN]; 193 u16 max_tx_vq; 194 u8 hash_key_length; 195 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE]; 196 }; 197 198 /* Control VQ buffers: protected by the rtnl lock */ 199 struct control_buf { 200 struct virtio_net_ctrl_hdr hdr; 201 virtio_net_ctrl_ack status; 202 struct virtio_net_ctrl_mq mq; 203 u8 promisc; 204 u8 allmulti; 205 __virtio16 vid; 206 __virtio64 offloads; 207 struct virtio_net_ctrl_rss rss; 208 }; 209 210 struct virtnet_info { 211 struct virtio_device *vdev; 212 struct virtqueue *cvq; 213 struct net_device *dev; 214 struct send_queue *sq; 215 struct receive_queue *rq; 216 unsigned int status; 217 218 /* Max # of queue pairs supported by the device */ 219 u16 max_queue_pairs; 220 221 /* # of queue pairs currently used by the driver */ 222 u16 curr_queue_pairs; 223 224 /* # of XDP queue pairs currently used by the driver */ 225 u16 xdp_queue_pairs; 226 227 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */ 228 bool xdp_enabled; 229 230 /* I like... big packets and I cannot lie! */ 231 bool big_packets; 232 233 /* number of sg entries allocated for big packets */ 234 unsigned int big_packets_num_skbfrags; 235 236 /* Host will merge rx buffers for big packets (shake it! shake it!) */ 237 bool mergeable_rx_bufs; 238 239 /* Host supports rss and/or hash report */ 240 bool has_rss; 241 bool has_rss_hash_report; 242 u8 rss_key_size; 243 u16 rss_indir_table_size; 244 u32 rss_hash_types_supported; 245 u32 rss_hash_types_saved; 246 247 /* Has control virtqueue */ 248 bool has_cvq; 249 250 /* Host can handle any s/g split between our header and packet data */ 251 bool any_header_sg; 252 253 /* Packet virtio header size */ 254 u8 hdr_len; 255 256 /* Work struct for delayed refilling if we run low on memory. */ 257 struct delayed_work refill; 258 259 /* Is delayed refill enabled? */ 260 bool refill_enabled; 261 262 /* The lock to synchronize the access to refill_enabled */ 263 spinlock_t refill_lock; 264 265 /* Work struct for config space updates */ 266 struct work_struct config_work; 267 268 /* Does the affinity hint is set for virtqueues? */ 269 bool affinity_hint_set; 270 271 /* CPU hotplug instances for online & dead */ 272 struct hlist_node node; 273 struct hlist_node node_dead; 274 275 struct control_buf *ctrl; 276 277 /* Ethtool settings */ 278 u8 duplex; 279 u32 speed; 280 281 /* Interrupt coalescing settings */ 282 u32 tx_usecs; 283 u32 rx_usecs; 284 u32 tx_max_packets; 285 u32 rx_max_packets; 286 287 unsigned long guest_offloads; 288 unsigned long guest_offloads_capable; 289 290 /* failover when STANDBY feature enabled */ 291 struct failover *failover; 292 }; 293 294 struct padded_vnet_hdr { 295 struct virtio_net_hdr_v1_hash hdr; 296 /* 297 * hdr is in a separate sg buffer, and data sg buffer shares same page 298 * with this header sg. This padding makes next sg 16 byte aligned 299 * after the header. 300 */ 301 char padding[12]; 302 }; 303 304 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf); 305 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf); 306 307 static bool is_xdp_frame(void *ptr) 308 { 309 return (unsigned long)ptr & VIRTIO_XDP_FLAG; 310 } 311 312 static void *xdp_to_ptr(struct xdp_frame *ptr) 313 { 314 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG); 315 } 316 317 static struct xdp_frame *ptr_to_xdp(void *ptr) 318 { 319 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG); 320 } 321 322 /* Converting between virtqueue no. and kernel tx/rx queue no. 323 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq 324 */ 325 static int vq2txq(struct virtqueue *vq) 326 { 327 return (vq->index - 1) / 2; 328 } 329 330 static int txq2vq(int txq) 331 { 332 return txq * 2 + 1; 333 } 334 335 static int vq2rxq(struct virtqueue *vq) 336 { 337 return vq->index / 2; 338 } 339 340 static int rxq2vq(int rxq) 341 { 342 return rxq * 2; 343 } 344 345 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb) 346 { 347 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb; 348 } 349 350 /* 351 * private is used to chain pages for big packets, put the whole 352 * most recent used list in the beginning for reuse 353 */ 354 static void give_pages(struct receive_queue *rq, struct page *page) 355 { 356 struct page *end; 357 358 /* Find end of list, sew whole thing into vi->rq.pages. */ 359 for (end = page; end->private; end = (struct page *)end->private); 360 end->private = (unsigned long)rq->pages; 361 rq->pages = page; 362 } 363 364 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask) 365 { 366 struct page *p = rq->pages; 367 368 if (p) { 369 rq->pages = (struct page *)p->private; 370 /* clear private here, it is used to chain pages */ 371 p->private = 0; 372 } else 373 p = alloc_page(gfp_mask); 374 return p; 375 } 376 377 static void enable_delayed_refill(struct virtnet_info *vi) 378 { 379 spin_lock_bh(&vi->refill_lock); 380 vi->refill_enabled = true; 381 spin_unlock_bh(&vi->refill_lock); 382 } 383 384 static void disable_delayed_refill(struct virtnet_info *vi) 385 { 386 spin_lock_bh(&vi->refill_lock); 387 vi->refill_enabled = false; 388 spin_unlock_bh(&vi->refill_lock); 389 } 390 391 static void virtqueue_napi_schedule(struct napi_struct *napi, 392 struct virtqueue *vq) 393 { 394 if (napi_schedule_prep(napi)) { 395 virtqueue_disable_cb(vq); 396 __napi_schedule(napi); 397 } 398 } 399 400 static void virtqueue_napi_complete(struct napi_struct *napi, 401 struct virtqueue *vq, int processed) 402 { 403 int opaque; 404 405 opaque = virtqueue_enable_cb_prepare(vq); 406 if (napi_complete_done(napi, processed)) { 407 if (unlikely(virtqueue_poll(vq, opaque))) 408 virtqueue_napi_schedule(napi, vq); 409 } else { 410 virtqueue_disable_cb(vq); 411 } 412 } 413 414 static void skb_xmit_done(struct virtqueue *vq) 415 { 416 struct virtnet_info *vi = vq->vdev->priv; 417 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi; 418 419 /* Suppress further interrupts. */ 420 virtqueue_disable_cb(vq); 421 422 if (napi->weight) 423 virtqueue_napi_schedule(napi, vq); 424 else 425 /* We were probably waiting for more output buffers. */ 426 netif_wake_subqueue(vi->dev, vq2txq(vq)); 427 } 428 429 #define MRG_CTX_HEADER_SHIFT 22 430 static void *mergeable_len_to_ctx(unsigned int truesize, 431 unsigned int headroom) 432 { 433 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize); 434 } 435 436 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx) 437 { 438 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT; 439 } 440 441 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx) 442 { 443 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1); 444 } 445 446 static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen, 447 unsigned int headroom, 448 unsigned int len) 449 { 450 struct sk_buff *skb; 451 452 skb = build_skb(buf, buflen); 453 if (unlikely(!skb)) 454 return NULL; 455 456 skb_reserve(skb, headroom); 457 skb_put(skb, len); 458 459 return skb; 460 } 461 462 /* Called from bottom half context */ 463 static struct sk_buff *page_to_skb(struct virtnet_info *vi, 464 struct receive_queue *rq, 465 struct page *page, unsigned int offset, 466 unsigned int len, unsigned int truesize, 467 unsigned int headroom) 468 { 469 struct sk_buff *skb; 470 struct virtio_net_hdr_mrg_rxbuf *hdr; 471 unsigned int copy, hdr_len, hdr_padded_len; 472 struct page *page_to_free = NULL; 473 int tailroom, shinfo_size; 474 char *p, *hdr_p, *buf; 475 476 p = page_address(page) + offset; 477 hdr_p = p; 478 479 hdr_len = vi->hdr_len; 480 if (vi->mergeable_rx_bufs) 481 hdr_padded_len = hdr_len; 482 else 483 hdr_padded_len = sizeof(struct padded_vnet_hdr); 484 485 buf = p - headroom; 486 len -= hdr_len; 487 offset += hdr_padded_len; 488 p += hdr_padded_len; 489 tailroom = truesize - headroom - hdr_padded_len - len; 490 491 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 492 493 /* copy small packet so we can reuse these pages */ 494 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) { 495 skb = virtnet_build_skb(buf, truesize, p - buf, len); 496 if (unlikely(!skb)) 497 return NULL; 498 499 page = (struct page *)page->private; 500 if (page) 501 give_pages(rq, page); 502 goto ok; 503 } 504 505 /* copy small packet so we can reuse these pages for small data */ 506 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN); 507 if (unlikely(!skb)) 508 return NULL; 509 510 /* Copy all frame if it fits skb->head, otherwise 511 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed. 512 */ 513 if (len <= skb_tailroom(skb)) 514 copy = len; 515 else 516 copy = ETH_HLEN; 517 skb_put_data(skb, p, copy); 518 519 len -= copy; 520 offset += copy; 521 522 if (vi->mergeable_rx_bufs) { 523 if (len) 524 skb_add_rx_frag(skb, 0, page, offset, len, truesize); 525 else 526 page_to_free = page; 527 goto ok; 528 } 529 530 /* 531 * Verify that we can indeed put this data into a skb. 532 * This is here to handle cases when the device erroneously 533 * tries to receive more than is possible. This is usually 534 * the case of a broken device. 535 */ 536 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) { 537 net_dbg_ratelimited("%s: too much data\n", skb->dev->name); 538 dev_kfree_skb(skb); 539 return NULL; 540 } 541 BUG_ON(offset >= PAGE_SIZE); 542 while (len) { 543 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len); 544 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset, 545 frag_size, truesize); 546 len -= frag_size; 547 page = (struct page *)page->private; 548 offset = 0; 549 } 550 551 if (page) 552 give_pages(rq, page); 553 554 ok: 555 hdr = skb_vnet_hdr(skb); 556 memcpy(hdr, hdr_p, hdr_len); 557 if (page_to_free) 558 put_page(page_to_free); 559 560 return skb; 561 } 562 563 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi) 564 { 565 unsigned int len; 566 unsigned int packets = 0; 567 unsigned int bytes = 0; 568 void *ptr; 569 570 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) { 571 if (likely(!is_xdp_frame(ptr))) { 572 struct sk_buff *skb = ptr; 573 574 pr_debug("Sent skb %p\n", skb); 575 576 bytes += skb->len; 577 napi_consume_skb(skb, in_napi); 578 } else { 579 struct xdp_frame *frame = ptr_to_xdp(ptr); 580 581 bytes += xdp_get_frame_len(frame); 582 xdp_return_frame(frame); 583 } 584 packets++; 585 } 586 587 /* Avoid overhead when no packets have been processed 588 * happens when called speculatively from start_xmit. 589 */ 590 if (!packets) 591 return; 592 593 u64_stats_update_begin(&sq->stats.syncp); 594 sq->stats.bytes += bytes; 595 sq->stats.packets += packets; 596 u64_stats_update_end(&sq->stats.syncp); 597 } 598 599 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q) 600 { 601 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs)) 602 return false; 603 else if (q < vi->curr_queue_pairs) 604 return true; 605 else 606 return false; 607 } 608 609 static void check_sq_full_and_disable(struct virtnet_info *vi, 610 struct net_device *dev, 611 struct send_queue *sq) 612 { 613 bool use_napi = sq->napi.weight; 614 int qnum; 615 616 qnum = sq - vi->sq; 617 618 /* If running out of space, stop queue to avoid getting packets that we 619 * are then unable to transmit. 620 * An alternative would be to force queuing layer to requeue the skb by 621 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be 622 * returned in a normal path of operation: it means that driver is not 623 * maintaining the TX queue stop/start state properly, and causes 624 * the stack to do a non-trivial amount of useless work. 625 * Since most packets only take 1 or 2 ring slots, stopping the queue 626 * early means 16 slots are typically wasted. 627 */ 628 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) { 629 netif_stop_subqueue(dev, qnum); 630 if (use_napi) { 631 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) 632 virtqueue_napi_schedule(&sq->napi, sq->vq); 633 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) { 634 /* More just got used, free them then recheck. */ 635 free_old_xmit_skbs(sq, false); 636 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) { 637 netif_start_subqueue(dev, qnum); 638 virtqueue_disable_cb(sq->vq); 639 } 640 } 641 } 642 } 643 644 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 645 struct send_queue *sq, 646 struct xdp_frame *xdpf) 647 { 648 struct virtio_net_hdr_mrg_rxbuf *hdr; 649 struct skb_shared_info *shinfo; 650 u8 nr_frags = 0; 651 int err, i; 652 653 if (unlikely(xdpf->headroom < vi->hdr_len)) 654 return -EOVERFLOW; 655 656 if (unlikely(xdp_frame_has_frags(xdpf))) { 657 shinfo = xdp_get_shared_info_from_frame(xdpf); 658 nr_frags = shinfo->nr_frags; 659 } 660 661 /* In wrapping function virtnet_xdp_xmit(), we need to free 662 * up the pending old buffers, where we need to calculate the 663 * position of skb_shared_info in xdp_get_frame_len() and 664 * xdp_return_frame(), which will involve to xdpf->data and 665 * xdpf->headroom. Therefore, we need to update the value of 666 * headroom synchronously here. 667 */ 668 xdpf->headroom -= vi->hdr_len; 669 xdpf->data -= vi->hdr_len; 670 /* Zero header and leave csum up to XDP layers */ 671 hdr = xdpf->data; 672 memset(hdr, 0, vi->hdr_len); 673 xdpf->len += vi->hdr_len; 674 675 sg_init_table(sq->sg, nr_frags + 1); 676 sg_set_buf(sq->sg, xdpf->data, xdpf->len); 677 for (i = 0; i < nr_frags; i++) { 678 skb_frag_t *frag = &shinfo->frags[i]; 679 680 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag), 681 skb_frag_size(frag), skb_frag_off(frag)); 682 } 683 684 err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1, 685 xdp_to_ptr(xdpf), GFP_ATOMIC); 686 if (unlikely(err)) 687 return -ENOSPC; /* Caller handle free/refcnt */ 688 689 return 0; 690 } 691 692 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on 693 * the current cpu, so it does not need to be locked. 694 * 695 * Here we use marco instead of inline functions because we have to deal with 696 * three issues at the same time: 1. the choice of sq. 2. judge and execute the 697 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline 698 * functions to perfectly solve these three problems at the same time. 699 */ 700 #define virtnet_xdp_get_sq(vi) ({ \ 701 int cpu = smp_processor_id(); \ 702 struct netdev_queue *txq; \ 703 typeof(vi) v = (vi); \ 704 unsigned int qp; \ 705 \ 706 if (v->curr_queue_pairs > nr_cpu_ids) { \ 707 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \ 708 qp += cpu; \ 709 txq = netdev_get_tx_queue(v->dev, qp); \ 710 __netif_tx_acquire(txq); \ 711 } else { \ 712 qp = cpu % v->curr_queue_pairs; \ 713 txq = netdev_get_tx_queue(v->dev, qp); \ 714 __netif_tx_lock(txq, cpu); \ 715 } \ 716 v->sq + qp; \ 717 }) 718 719 #define virtnet_xdp_put_sq(vi, q) { \ 720 struct netdev_queue *txq; \ 721 typeof(vi) v = (vi); \ 722 \ 723 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \ 724 if (v->curr_queue_pairs > nr_cpu_ids) \ 725 __netif_tx_release(txq); \ 726 else \ 727 __netif_tx_unlock(txq); \ 728 } 729 730 static int virtnet_xdp_xmit(struct net_device *dev, 731 int n, struct xdp_frame **frames, u32 flags) 732 { 733 struct virtnet_info *vi = netdev_priv(dev); 734 struct receive_queue *rq = vi->rq; 735 struct bpf_prog *xdp_prog; 736 struct send_queue *sq; 737 unsigned int len; 738 int packets = 0; 739 int bytes = 0; 740 int nxmit = 0; 741 int kicks = 0; 742 void *ptr; 743 int ret; 744 int i; 745 746 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this 747 * indicate XDP resources have been successfully allocated. 748 */ 749 xdp_prog = rcu_access_pointer(rq->xdp_prog); 750 if (!xdp_prog) 751 return -ENXIO; 752 753 sq = virtnet_xdp_get_sq(vi); 754 755 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) { 756 ret = -EINVAL; 757 goto out; 758 } 759 760 /* Free up any pending old buffers before queueing new ones. */ 761 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) { 762 if (likely(is_xdp_frame(ptr))) { 763 struct xdp_frame *frame = ptr_to_xdp(ptr); 764 765 bytes += xdp_get_frame_len(frame); 766 xdp_return_frame(frame); 767 } else { 768 struct sk_buff *skb = ptr; 769 770 bytes += skb->len; 771 napi_consume_skb(skb, false); 772 } 773 packets++; 774 } 775 776 for (i = 0; i < n; i++) { 777 struct xdp_frame *xdpf = frames[i]; 778 779 if (__virtnet_xdp_xmit_one(vi, sq, xdpf)) 780 break; 781 nxmit++; 782 } 783 ret = nxmit; 784 785 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq)) 786 check_sq_full_and_disable(vi, dev, sq); 787 788 if (flags & XDP_XMIT_FLUSH) { 789 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) 790 kicks = 1; 791 } 792 out: 793 u64_stats_update_begin(&sq->stats.syncp); 794 sq->stats.bytes += bytes; 795 sq->stats.packets += packets; 796 sq->stats.xdp_tx += n; 797 sq->stats.xdp_tx_drops += n - nxmit; 798 sq->stats.kicks += kicks; 799 u64_stats_update_end(&sq->stats.syncp); 800 801 virtnet_xdp_put_sq(vi, sq); 802 return ret; 803 } 804 805 static void put_xdp_frags(struct xdp_buff *xdp) 806 { 807 struct skb_shared_info *shinfo; 808 struct page *xdp_page; 809 int i; 810 811 if (xdp_buff_has_frags(xdp)) { 812 shinfo = xdp_get_shared_info_from_buff(xdp); 813 for (i = 0; i < shinfo->nr_frags; i++) { 814 xdp_page = skb_frag_page(&shinfo->frags[i]); 815 put_page(xdp_page); 816 } 817 } 818 } 819 820 static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp, 821 struct net_device *dev, 822 unsigned int *xdp_xmit, 823 struct virtnet_rq_stats *stats) 824 { 825 struct xdp_frame *xdpf; 826 int err; 827 u32 act; 828 829 act = bpf_prog_run_xdp(xdp_prog, xdp); 830 stats->xdp_packets++; 831 832 switch (act) { 833 case XDP_PASS: 834 return act; 835 836 case XDP_TX: 837 stats->xdp_tx++; 838 xdpf = xdp_convert_buff_to_frame(xdp); 839 if (unlikely(!xdpf)) { 840 netdev_dbg(dev, "convert buff to frame failed for xdp\n"); 841 return XDP_DROP; 842 } 843 844 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0); 845 if (unlikely(!err)) { 846 xdp_return_frame_rx_napi(xdpf); 847 } else if (unlikely(err < 0)) { 848 trace_xdp_exception(dev, xdp_prog, act); 849 return XDP_DROP; 850 } 851 *xdp_xmit |= VIRTIO_XDP_TX; 852 return act; 853 854 case XDP_REDIRECT: 855 stats->xdp_redirects++; 856 err = xdp_do_redirect(dev, xdp, xdp_prog); 857 if (err) 858 return XDP_DROP; 859 860 *xdp_xmit |= VIRTIO_XDP_REDIR; 861 return act; 862 863 default: 864 bpf_warn_invalid_xdp_action(dev, xdp_prog, act); 865 fallthrough; 866 case XDP_ABORTED: 867 trace_xdp_exception(dev, xdp_prog, act); 868 fallthrough; 869 case XDP_DROP: 870 return XDP_DROP; 871 } 872 } 873 874 static unsigned int virtnet_get_headroom(struct virtnet_info *vi) 875 { 876 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0; 877 } 878 879 /* We copy the packet for XDP in the following cases: 880 * 881 * 1) Packet is scattered across multiple rx buffers. 882 * 2) Headroom space is insufficient. 883 * 884 * This is inefficient but it's a temporary condition that 885 * we hit right after XDP is enabled and until queue is refilled 886 * with large buffers with sufficient headroom - so it should affect 887 * at most queue size packets. 888 * Afterwards, the conditions to enable 889 * XDP should preclude the underlying device from sending packets 890 * across multiple buffers (num_buf > 1), and we make sure buffers 891 * have enough headroom. 892 */ 893 static struct page *xdp_linearize_page(struct receive_queue *rq, 894 int *num_buf, 895 struct page *p, 896 int offset, 897 int page_off, 898 unsigned int *len) 899 { 900 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 901 struct page *page; 902 903 if (page_off + *len + tailroom > PAGE_SIZE) 904 return NULL; 905 906 page = alloc_page(GFP_ATOMIC); 907 if (!page) 908 return NULL; 909 910 memcpy(page_address(page) + page_off, page_address(p) + offset, *len); 911 page_off += *len; 912 913 while (--*num_buf) { 914 unsigned int buflen; 915 void *buf; 916 int off; 917 918 buf = virtqueue_get_buf(rq->vq, &buflen); 919 if (unlikely(!buf)) 920 goto err_buf; 921 922 p = virt_to_head_page(buf); 923 off = buf - page_address(p); 924 925 /* guard against a misconfigured or uncooperative backend that 926 * is sending packet larger than the MTU. 927 */ 928 if ((page_off + buflen + tailroom) > PAGE_SIZE) { 929 put_page(p); 930 goto err_buf; 931 } 932 933 memcpy(page_address(page) + page_off, 934 page_address(p) + off, buflen); 935 page_off += buflen; 936 put_page(p); 937 } 938 939 /* Headroom does not contribute to packet length */ 940 *len = page_off - VIRTIO_XDP_HEADROOM; 941 return page; 942 err_buf: 943 __free_pages(page, 0); 944 return NULL; 945 } 946 947 static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi, 948 unsigned int xdp_headroom, 949 void *buf, 950 unsigned int len) 951 { 952 unsigned int header_offset; 953 unsigned int headroom; 954 unsigned int buflen; 955 struct sk_buff *skb; 956 957 header_offset = VIRTNET_RX_PAD + xdp_headroom; 958 headroom = vi->hdr_len + header_offset; 959 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 960 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 961 962 skb = virtnet_build_skb(buf, buflen, headroom, len); 963 if (unlikely(!skb)) 964 return NULL; 965 966 buf += header_offset; 967 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len); 968 969 return skb; 970 } 971 972 static struct sk_buff *receive_small_xdp(struct net_device *dev, 973 struct virtnet_info *vi, 974 struct receive_queue *rq, 975 struct bpf_prog *xdp_prog, 976 void *buf, 977 unsigned int xdp_headroom, 978 unsigned int len, 979 unsigned int *xdp_xmit, 980 struct virtnet_rq_stats *stats) 981 { 982 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom; 983 unsigned int headroom = vi->hdr_len + header_offset; 984 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset; 985 struct page *page = virt_to_head_page(buf); 986 struct page *xdp_page; 987 unsigned int buflen; 988 struct xdp_buff xdp; 989 struct sk_buff *skb; 990 unsigned int metasize = 0; 991 u32 act; 992 993 if (unlikely(hdr->hdr.gso_type)) 994 goto err_xdp; 995 996 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 997 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 998 999 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) { 1000 int offset = buf - page_address(page) + header_offset; 1001 unsigned int tlen = len + vi->hdr_len; 1002 int num_buf = 1; 1003 1004 xdp_headroom = virtnet_get_headroom(vi); 1005 header_offset = VIRTNET_RX_PAD + xdp_headroom; 1006 headroom = vi->hdr_len + header_offset; 1007 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) + 1008 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1009 xdp_page = xdp_linearize_page(rq, &num_buf, page, 1010 offset, header_offset, 1011 &tlen); 1012 if (!xdp_page) 1013 goto err_xdp; 1014 1015 buf = page_address(xdp_page); 1016 put_page(page); 1017 page = xdp_page; 1018 } 1019 1020 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq); 1021 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len, 1022 xdp_headroom, len, true); 1023 1024 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats); 1025 1026 switch (act) { 1027 case XDP_PASS: 1028 /* Recalculate length in case bpf program changed it */ 1029 len = xdp.data_end - xdp.data; 1030 metasize = xdp.data - xdp.data_meta; 1031 break; 1032 1033 case XDP_TX: 1034 case XDP_REDIRECT: 1035 goto xdp_xmit; 1036 1037 default: 1038 goto err_xdp; 1039 } 1040 1041 skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len); 1042 if (unlikely(!skb)) 1043 goto err; 1044 1045 if (metasize) 1046 skb_metadata_set(skb, metasize); 1047 1048 return skb; 1049 1050 err_xdp: 1051 stats->xdp_drops++; 1052 err: 1053 stats->drops++; 1054 put_page(page); 1055 xdp_xmit: 1056 return NULL; 1057 } 1058 1059 static struct sk_buff *receive_small(struct net_device *dev, 1060 struct virtnet_info *vi, 1061 struct receive_queue *rq, 1062 void *buf, void *ctx, 1063 unsigned int len, 1064 unsigned int *xdp_xmit, 1065 struct virtnet_rq_stats *stats) 1066 { 1067 unsigned int xdp_headroom = (unsigned long)ctx; 1068 struct page *page = virt_to_head_page(buf); 1069 struct sk_buff *skb; 1070 1071 len -= vi->hdr_len; 1072 stats->bytes += len; 1073 1074 if (unlikely(len > GOOD_PACKET_LEN)) { 1075 pr_debug("%s: rx error: len %u exceeds max size %d\n", 1076 dev->name, len, GOOD_PACKET_LEN); 1077 dev->stats.rx_length_errors++; 1078 goto err; 1079 } 1080 1081 if (unlikely(vi->xdp_enabled)) { 1082 struct bpf_prog *xdp_prog; 1083 1084 rcu_read_lock(); 1085 xdp_prog = rcu_dereference(rq->xdp_prog); 1086 if (xdp_prog) { 1087 skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf, 1088 xdp_headroom, len, xdp_xmit, 1089 stats); 1090 rcu_read_unlock(); 1091 return skb; 1092 } 1093 rcu_read_unlock(); 1094 } 1095 1096 skb = receive_small_build_skb(vi, xdp_headroom, buf, len); 1097 if (likely(skb)) 1098 return skb; 1099 1100 err: 1101 stats->drops++; 1102 put_page(page); 1103 return NULL; 1104 } 1105 1106 static struct sk_buff *receive_big(struct net_device *dev, 1107 struct virtnet_info *vi, 1108 struct receive_queue *rq, 1109 void *buf, 1110 unsigned int len, 1111 struct virtnet_rq_stats *stats) 1112 { 1113 struct page *page = buf; 1114 struct sk_buff *skb = 1115 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0); 1116 1117 stats->bytes += len - vi->hdr_len; 1118 if (unlikely(!skb)) 1119 goto err; 1120 1121 return skb; 1122 1123 err: 1124 stats->drops++; 1125 give_pages(rq, page); 1126 return NULL; 1127 } 1128 1129 static void mergeable_buf_free(struct receive_queue *rq, int num_buf, 1130 struct net_device *dev, 1131 struct virtnet_rq_stats *stats) 1132 { 1133 struct page *page; 1134 void *buf; 1135 int len; 1136 1137 while (num_buf-- > 1) { 1138 buf = virtqueue_get_buf(rq->vq, &len); 1139 if (unlikely(!buf)) { 1140 pr_debug("%s: rx error: %d buffers missing\n", 1141 dev->name, num_buf); 1142 dev->stats.rx_length_errors++; 1143 break; 1144 } 1145 stats->bytes += len; 1146 page = virt_to_head_page(buf); 1147 put_page(page); 1148 } 1149 } 1150 1151 /* Why not use xdp_build_skb_from_frame() ? 1152 * XDP core assumes that xdp frags are PAGE_SIZE in length, while in 1153 * virtio-net there are 2 points that do not match its requirements: 1154 * 1. The size of the prefilled buffer is not fixed before xdp is set. 1155 * 2. xdp_build_skb_from_frame() does more checks that we don't need, 1156 * like eth_type_trans() (which virtio-net does in receive_buf()). 1157 */ 1158 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev, 1159 struct virtnet_info *vi, 1160 struct xdp_buff *xdp, 1161 unsigned int xdp_frags_truesz) 1162 { 1163 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp); 1164 unsigned int headroom, data_len; 1165 struct sk_buff *skb; 1166 int metasize; 1167 u8 nr_frags; 1168 1169 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) { 1170 pr_debug("Error building skb as missing reserved tailroom for xdp"); 1171 return NULL; 1172 } 1173 1174 if (unlikely(xdp_buff_has_frags(xdp))) 1175 nr_frags = sinfo->nr_frags; 1176 1177 skb = build_skb(xdp->data_hard_start, xdp->frame_sz); 1178 if (unlikely(!skb)) 1179 return NULL; 1180 1181 headroom = xdp->data - xdp->data_hard_start; 1182 data_len = xdp->data_end - xdp->data; 1183 skb_reserve(skb, headroom); 1184 __skb_put(skb, data_len); 1185 1186 metasize = xdp->data - xdp->data_meta; 1187 metasize = metasize > 0 ? metasize : 0; 1188 if (metasize) 1189 skb_metadata_set(skb, metasize); 1190 1191 if (unlikely(xdp_buff_has_frags(xdp))) 1192 xdp_update_skb_shared_info(skb, nr_frags, 1193 sinfo->xdp_frags_size, 1194 xdp_frags_truesz, 1195 xdp_buff_is_frag_pfmemalloc(xdp)); 1196 1197 return skb; 1198 } 1199 1200 /* TODO: build xdp in big mode */ 1201 static int virtnet_build_xdp_buff_mrg(struct net_device *dev, 1202 struct virtnet_info *vi, 1203 struct receive_queue *rq, 1204 struct xdp_buff *xdp, 1205 void *buf, 1206 unsigned int len, 1207 unsigned int frame_sz, 1208 int *num_buf, 1209 unsigned int *xdp_frags_truesize, 1210 struct virtnet_rq_stats *stats) 1211 { 1212 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 1213 unsigned int headroom, tailroom, room; 1214 unsigned int truesize, cur_frag_size; 1215 struct skb_shared_info *shinfo; 1216 unsigned int xdp_frags_truesz = 0; 1217 struct page *page; 1218 skb_frag_t *frag; 1219 int offset; 1220 void *ctx; 1221 1222 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq); 1223 xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM, 1224 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true); 1225 1226 if (!*num_buf) 1227 return 0; 1228 1229 if (*num_buf > 1) { 1230 /* If we want to build multi-buffer xdp, we need 1231 * to specify that the flags of xdp_buff have the 1232 * XDP_FLAGS_HAS_FRAG bit. 1233 */ 1234 if (!xdp_buff_has_frags(xdp)) 1235 xdp_buff_set_frags_flag(xdp); 1236 1237 shinfo = xdp_get_shared_info_from_buff(xdp); 1238 shinfo->nr_frags = 0; 1239 shinfo->xdp_frags_size = 0; 1240 } 1241 1242 if (*num_buf > MAX_SKB_FRAGS + 1) 1243 return -EINVAL; 1244 1245 while (--*num_buf > 0) { 1246 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 1247 if (unlikely(!buf)) { 1248 pr_debug("%s: rx error: %d buffers out of %d missing\n", 1249 dev->name, *num_buf, 1250 virtio16_to_cpu(vi->vdev, hdr->num_buffers)); 1251 dev->stats.rx_length_errors++; 1252 goto err; 1253 } 1254 1255 stats->bytes += len; 1256 page = virt_to_head_page(buf); 1257 offset = buf - page_address(page); 1258 1259 truesize = mergeable_ctx_to_truesize(ctx); 1260 headroom = mergeable_ctx_to_headroom(ctx); 1261 tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1262 room = SKB_DATA_ALIGN(headroom + tailroom); 1263 1264 cur_frag_size = truesize; 1265 xdp_frags_truesz += cur_frag_size; 1266 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) { 1267 put_page(page); 1268 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1269 dev->name, len, (unsigned long)(truesize - room)); 1270 dev->stats.rx_length_errors++; 1271 goto err; 1272 } 1273 1274 frag = &shinfo->frags[shinfo->nr_frags++]; 1275 skb_frag_fill_page_desc(frag, page, offset, len); 1276 if (page_is_pfmemalloc(page)) 1277 xdp_buff_set_frag_pfmemalloc(xdp); 1278 1279 shinfo->xdp_frags_size += len; 1280 } 1281 1282 *xdp_frags_truesize = xdp_frags_truesz; 1283 return 0; 1284 1285 err: 1286 put_xdp_frags(xdp); 1287 return -EINVAL; 1288 } 1289 1290 static void *mergeable_xdp_get_buf(struct virtnet_info *vi, 1291 struct receive_queue *rq, 1292 struct bpf_prog *xdp_prog, 1293 void *ctx, 1294 unsigned int *frame_sz, 1295 int *num_buf, 1296 struct page **page, 1297 int offset, 1298 unsigned int *len, 1299 struct virtio_net_hdr_mrg_rxbuf *hdr) 1300 { 1301 unsigned int truesize = mergeable_ctx_to_truesize(ctx); 1302 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 1303 struct page *xdp_page; 1304 unsigned int xdp_room; 1305 1306 /* Transient failure which in theory could occur if 1307 * in-flight packets from before XDP was enabled reach 1308 * the receive path after XDP is loaded. 1309 */ 1310 if (unlikely(hdr->hdr.gso_type)) 1311 return NULL; 1312 1313 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers 1314 * with headroom may add hole in truesize, which 1315 * make their length exceed PAGE_SIZE. So we disabled the 1316 * hole mechanism for xdp. See add_recvbuf_mergeable(). 1317 */ 1318 *frame_sz = truesize; 1319 1320 if (likely(headroom >= virtnet_get_headroom(vi) && 1321 (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) { 1322 return page_address(*page) + offset; 1323 } 1324 1325 /* This happens when headroom is not enough because 1326 * of the buffer was prefilled before XDP is set. 1327 * This should only happen for the first several packets. 1328 * In fact, vq reset can be used here to help us clean up 1329 * the prefilled buffers, but many existing devices do not 1330 * support it, and we don't want to bother users who are 1331 * using xdp normally. 1332 */ 1333 if (!xdp_prog->aux->xdp_has_frags) { 1334 /* linearize data for XDP */ 1335 xdp_page = xdp_linearize_page(rq, num_buf, 1336 *page, offset, 1337 VIRTIO_XDP_HEADROOM, 1338 len); 1339 if (!xdp_page) 1340 return NULL; 1341 } else { 1342 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM + 1343 sizeof(struct skb_shared_info)); 1344 if (*len + xdp_room > PAGE_SIZE) 1345 return NULL; 1346 1347 xdp_page = alloc_page(GFP_ATOMIC); 1348 if (!xdp_page) 1349 return NULL; 1350 1351 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM, 1352 page_address(*page) + offset, *len); 1353 } 1354 1355 *frame_sz = PAGE_SIZE; 1356 1357 put_page(*page); 1358 1359 *page = xdp_page; 1360 1361 return page_address(*page) + VIRTIO_XDP_HEADROOM; 1362 } 1363 1364 static struct sk_buff *receive_mergeable_xdp(struct net_device *dev, 1365 struct virtnet_info *vi, 1366 struct receive_queue *rq, 1367 struct bpf_prog *xdp_prog, 1368 void *buf, 1369 void *ctx, 1370 unsigned int len, 1371 unsigned int *xdp_xmit, 1372 struct virtnet_rq_stats *stats) 1373 { 1374 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 1375 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 1376 struct page *page = virt_to_head_page(buf); 1377 int offset = buf - page_address(page); 1378 unsigned int xdp_frags_truesz = 0; 1379 struct sk_buff *head_skb; 1380 unsigned int frame_sz; 1381 struct xdp_buff xdp; 1382 void *data; 1383 u32 act; 1384 int err; 1385 1386 data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page, 1387 offset, &len, hdr); 1388 if (unlikely(!data)) 1389 goto err_xdp; 1390 1391 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz, 1392 &num_buf, &xdp_frags_truesz, stats); 1393 if (unlikely(err)) 1394 goto err_xdp; 1395 1396 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats); 1397 1398 switch (act) { 1399 case XDP_PASS: 1400 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz); 1401 if (unlikely(!head_skb)) 1402 break; 1403 return head_skb; 1404 1405 case XDP_TX: 1406 case XDP_REDIRECT: 1407 return NULL; 1408 1409 default: 1410 break; 1411 } 1412 1413 put_xdp_frags(&xdp); 1414 1415 err_xdp: 1416 put_page(page); 1417 mergeable_buf_free(rq, num_buf, dev, stats); 1418 1419 stats->xdp_drops++; 1420 stats->drops++; 1421 return NULL; 1422 } 1423 1424 static struct sk_buff *receive_mergeable(struct net_device *dev, 1425 struct virtnet_info *vi, 1426 struct receive_queue *rq, 1427 void *buf, 1428 void *ctx, 1429 unsigned int len, 1430 unsigned int *xdp_xmit, 1431 struct virtnet_rq_stats *stats) 1432 { 1433 struct virtio_net_hdr_mrg_rxbuf *hdr = buf; 1434 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers); 1435 struct page *page = virt_to_head_page(buf); 1436 int offset = buf - page_address(page); 1437 struct sk_buff *head_skb, *curr_skb; 1438 unsigned int truesize = mergeable_ctx_to_truesize(ctx); 1439 unsigned int headroom = mergeable_ctx_to_headroom(ctx); 1440 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1441 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1442 1443 head_skb = NULL; 1444 stats->bytes += len - vi->hdr_len; 1445 1446 if (unlikely(len > truesize - room)) { 1447 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1448 dev->name, len, (unsigned long)(truesize - room)); 1449 dev->stats.rx_length_errors++; 1450 goto err_skb; 1451 } 1452 1453 if (unlikely(vi->xdp_enabled)) { 1454 struct bpf_prog *xdp_prog; 1455 1456 rcu_read_lock(); 1457 xdp_prog = rcu_dereference(rq->xdp_prog); 1458 if (xdp_prog) { 1459 head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx, 1460 len, xdp_xmit, stats); 1461 rcu_read_unlock(); 1462 return head_skb; 1463 } 1464 rcu_read_unlock(); 1465 } 1466 1467 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom); 1468 curr_skb = head_skb; 1469 1470 if (unlikely(!curr_skb)) 1471 goto err_skb; 1472 while (--num_buf) { 1473 int num_skb_frags; 1474 1475 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx); 1476 if (unlikely(!buf)) { 1477 pr_debug("%s: rx error: %d buffers out of %d missing\n", 1478 dev->name, num_buf, 1479 virtio16_to_cpu(vi->vdev, 1480 hdr->num_buffers)); 1481 dev->stats.rx_length_errors++; 1482 goto err_buf; 1483 } 1484 1485 stats->bytes += len; 1486 page = virt_to_head_page(buf); 1487 1488 truesize = mergeable_ctx_to_truesize(ctx); 1489 headroom = mergeable_ctx_to_headroom(ctx); 1490 tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1491 room = SKB_DATA_ALIGN(headroom + tailroom); 1492 if (unlikely(len > truesize - room)) { 1493 pr_debug("%s: rx error: len %u exceeds truesize %lu\n", 1494 dev->name, len, (unsigned long)(truesize - room)); 1495 dev->stats.rx_length_errors++; 1496 goto err_skb; 1497 } 1498 1499 num_skb_frags = skb_shinfo(curr_skb)->nr_frags; 1500 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) { 1501 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC); 1502 1503 if (unlikely(!nskb)) 1504 goto err_skb; 1505 if (curr_skb == head_skb) 1506 skb_shinfo(curr_skb)->frag_list = nskb; 1507 else 1508 curr_skb->next = nskb; 1509 curr_skb = nskb; 1510 head_skb->truesize += nskb->truesize; 1511 num_skb_frags = 0; 1512 } 1513 if (curr_skb != head_skb) { 1514 head_skb->data_len += len; 1515 head_skb->len += len; 1516 head_skb->truesize += truesize; 1517 } 1518 offset = buf - page_address(page); 1519 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) { 1520 put_page(page); 1521 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1, 1522 len, truesize); 1523 } else { 1524 skb_add_rx_frag(curr_skb, num_skb_frags, page, 1525 offset, len, truesize); 1526 } 1527 } 1528 1529 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len); 1530 return head_skb; 1531 1532 err_skb: 1533 put_page(page); 1534 mergeable_buf_free(rq, num_buf, dev, stats); 1535 1536 err_buf: 1537 stats->drops++; 1538 dev_kfree_skb(head_skb); 1539 return NULL; 1540 } 1541 1542 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash, 1543 struct sk_buff *skb) 1544 { 1545 enum pkt_hash_types rss_hash_type; 1546 1547 if (!hdr_hash || !skb) 1548 return; 1549 1550 switch (__le16_to_cpu(hdr_hash->hash_report)) { 1551 case VIRTIO_NET_HASH_REPORT_TCPv4: 1552 case VIRTIO_NET_HASH_REPORT_UDPv4: 1553 case VIRTIO_NET_HASH_REPORT_TCPv6: 1554 case VIRTIO_NET_HASH_REPORT_UDPv6: 1555 case VIRTIO_NET_HASH_REPORT_TCPv6_EX: 1556 case VIRTIO_NET_HASH_REPORT_UDPv6_EX: 1557 rss_hash_type = PKT_HASH_TYPE_L4; 1558 break; 1559 case VIRTIO_NET_HASH_REPORT_IPv4: 1560 case VIRTIO_NET_HASH_REPORT_IPv6: 1561 case VIRTIO_NET_HASH_REPORT_IPv6_EX: 1562 rss_hash_type = PKT_HASH_TYPE_L3; 1563 break; 1564 case VIRTIO_NET_HASH_REPORT_NONE: 1565 default: 1566 rss_hash_type = PKT_HASH_TYPE_NONE; 1567 } 1568 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type); 1569 } 1570 1571 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq, 1572 void *buf, unsigned int len, void **ctx, 1573 unsigned int *xdp_xmit, 1574 struct virtnet_rq_stats *stats) 1575 { 1576 struct net_device *dev = vi->dev; 1577 struct sk_buff *skb; 1578 struct virtio_net_hdr_mrg_rxbuf *hdr; 1579 1580 if (unlikely(len < vi->hdr_len + ETH_HLEN)) { 1581 pr_debug("%s: short packet %i\n", dev->name, len); 1582 dev->stats.rx_length_errors++; 1583 virtnet_rq_free_unused_buf(rq->vq, buf); 1584 return; 1585 } 1586 1587 if (vi->mergeable_rx_bufs) 1588 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit, 1589 stats); 1590 else if (vi->big_packets) 1591 skb = receive_big(dev, vi, rq, buf, len, stats); 1592 else 1593 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats); 1594 1595 if (unlikely(!skb)) 1596 return; 1597 1598 hdr = skb_vnet_hdr(skb); 1599 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report) 1600 virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb); 1601 1602 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) 1603 skb->ip_summed = CHECKSUM_UNNECESSARY; 1604 1605 if (virtio_net_hdr_to_skb(skb, &hdr->hdr, 1606 virtio_is_little_endian(vi->vdev))) { 1607 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n", 1608 dev->name, hdr->hdr.gso_type, 1609 hdr->hdr.gso_size); 1610 goto frame_err; 1611 } 1612 1613 skb_record_rx_queue(skb, vq2rxq(rq->vq)); 1614 skb->protocol = eth_type_trans(skb, dev); 1615 pr_debug("Receiving skb proto 0x%04x len %i type %i\n", 1616 ntohs(skb->protocol), skb->len, skb->pkt_type); 1617 1618 napi_gro_receive(&rq->napi, skb); 1619 return; 1620 1621 frame_err: 1622 dev->stats.rx_frame_errors++; 1623 dev_kfree_skb(skb); 1624 } 1625 1626 /* Unlike mergeable buffers, all buffers are allocated to the 1627 * same size, except for the headroom. For this reason we do 1628 * not need to use mergeable_len_to_ctx here - it is enough 1629 * to store the headroom as the context ignoring the truesize. 1630 */ 1631 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq, 1632 gfp_t gfp) 1633 { 1634 struct page_frag *alloc_frag = &rq->alloc_frag; 1635 char *buf; 1636 unsigned int xdp_headroom = virtnet_get_headroom(vi); 1637 void *ctx = (void *)(unsigned long)xdp_headroom; 1638 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom; 1639 int err; 1640 1641 len = SKB_DATA_ALIGN(len) + 1642 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 1643 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp))) 1644 return -ENOMEM; 1645 1646 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1647 get_page(alloc_frag->page); 1648 alloc_frag->offset += len; 1649 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom, 1650 vi->hdr_len + GOOD_PACKET_LEN); 1651 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1652 if (err < 0) 1653 put_page(virt_to_head_page(buf)); 1654 return err; 1655 } 1656 1657 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq, 1658 gfp_t gfp) 1659 { 1660 struct page *first, *list = NULL; 1661 char *p; 1662 int i, err, offset; 1663 1664 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2); 1665 1666 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */ 1667 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) { 1668 first = get_a_page(rq, gfp); 1669 if (!first) { 1670 if (list) 1671 give_pages(rq, list); 1672 return -ENOMEM; 1673 } 1674 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE); 1675 1676 /* chain new page in list head to match sg */ 1677 first->private = (unsigned long)list; 1678 list = first; 1679 } 1680 1681 first = get_a_page(rq, gfp); 1682 if (!first) { 1683 give_pages(rq, list); 1684 return -ENOMEM; 1685 } 1686 p = page_address(first); 1687 1688 /* rq->sg[0], rq->sg[1] share the same page */ 1689 /* a separated rq->sg[0] for header - required in case !any_header_sg */ 1690 sg_set_buf(&rq->sg[0], p, vi->hdr_len); 1691 1692 /* rq->sg[1] for data packet, from offset */ 1693 offset = sizeof(struct padded_vnet_hdr); 1694 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset); 1695 1696 /* chain first in list head */ 1697 first->private = (unsigned long)list; 1698 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2, 1699 first, gfp); 1700 if (err < 0) 1701 give_pages(rq, first); 1702 1703 return err; 1704 } 1705 1706 static unsigned int get_mergeable_buf_len(struct receive_queue *rq, 1707 struct ewma_pkt_len *avg_pkt_len, 1708 unsigned int room) 1709 { 1710 struct virtnet_info *vi = rq->vq->vdev->priv; 1711 const size_t hdr_len = vi->hdr_len; 1712 unsigned int len; 1713 1714 if (room) 1715 return PAGE_SIZE - room; 1716 1717 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len), 1718 rq->min_buf_len, PAGE_SIZE - hdr_len); 1719 1720 return ALIGN(len, L1_CACHE_BYTES); 1721 } 1722 1723 static int add_recvbuf_mergeable(struct virtnet_info *vi, 1724 struct receive_queue *rq, gfp_t gfp) 1725 { 1726 struct page_frag *alloc_frag = &rq->alloc_frag; 1727 unsigned int headroom = virtnet_get_headroom(vi); 1728 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 1729 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom); 1730 char *buf; 1731 void *ctx; 1732 int err; 1733 unsigned int len, hole; 1734 1735 /* Extra tailroom is needed to satisfy XDP's assumption. This 1736 * means rx frags coalescing won't work, but consider we've 1737 * disabled GSO for XDP, it won't be a big issue. 1738 */ 1739 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room); 1740 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp))) 1741 return -ENOMEM; 1742 1743 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset; 1744 buf += headroom; /* advance address leaving hole at front of pkt */ 1745 get_page(alloc_frag->page); 1746 alloc_frag->offset += len + room; 1747 hole = alloc_frag->size - alloc_frag->offset; 1748 if (hole < len + room) { 1749 /* To avoid internal fragmentation, if there is very likely not 1750 * enough space for another buffer, add the remaining space to 1751 * the current buffer. 1752 * XDP core assumes that frame_size of xdp_buff and the length 1753 * of the frag are PAGE_SIZE, so we disable the hole mechanism. 1754 */ 1755 if (!headroom) 1756 len += hole; 1757 alloc_frag->offset += hole; 1758 } 1759 1760 sg_init_one(rq->sg, buf, len); 1761 ctx = mergeable_len_to_ctx(len + room, headroom); 1762 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp); 1763 if (err < 0) 1764 put_page(virt_to_head_page(buf)); 1765 1766 return err; 1767 } 1768 1769 /* 1770 * Returns false if we couldn't fill entirely (OOM). 1771 * 1772 * Normally run in the receive path, but can also be run from ndo_open 1773 * before we're receiving packets, or from refill_work which is 1774 * careful to disable receiving (using napi_disable). 1775 */ 1776 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq, 1777 gfp_t gfp) 1778 { 1779 int err; 1780 bool oom; 1781 1782 do { 1783 if (vi->mergeable_rx_bufs) 1784 err = add_recvbuf_mergeable(vi, rq, gfp); 1785 else if (vi->big_packets) 1786 err = add_recvbuf_big(vi, rq, gfp); 1787 else 1788 err = add_recvbuf_small(vi, rq, gfp); 1789 1790 oom = err == -ENOMEM; 1791 if (err) 1792 break; 1793 } while (rq->vq->num_free); 1794 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) { 1795 unsigned long flags; 1796 1797 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp); 1798 rq->stats.kicks++; 1799 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags); 1800 } 1801 1802 return !oom; 1803 } 1804 1805 static void skb_recv_done(struct virtqueue *rvq) 1806 { 1807 struct virtnet_info *vi = rvq->vdev->priv; 1808 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)]; 1809 1810 virtqueue_napi_schedule(&rq->napi, rvq); 1811 } 1812 1813 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi) 1814 { 1815 napi_enable(napi); 1816 1817 /* If all buffers were filled by other side before we napi_enabled, we 1818 * won't get another interrupt, so process any outstanding packets now. 1819 * Call local_bh_enable after to trigger softIRQ processing. 1820 */ 1821 local_bh_disable(); 1822 virtqueue_napi_schedule(napi, vq); 1823 local_bh_enable(); 1824 } 1825 1826 static void virtnet_napi_tx_enable(struct virtnet_info *vi, 1827 struct virtqueue *vq, 1828 struct napi_struct *napi) 1829 { 1830 if (!napi->weight) 1831 return; 1832 1833 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only 1834 * enable the feature if this is likely affine with the transmit path. 1835 */ 1836 if (!vi->affinity_hint_set) { 1837 napi->weight = 0; 1838 return; 1839 } 1840 1841 return virtnet_napi_enable(vq, napi); 1842 } 1843 1844 static void virtnet_napi_tx_disable(struct napi_struct *napi) 1845 { 1846 if (napi->weight) 1847 napi_disable(napi); 1848 } 1849 1850 static void refill_work(struct work_struct *work) 1851 { 1852 struct virtnet_info *vi = 1853 container_of(work, struct virtnet_info, refill.work); 1854 bool still_empty; 1855 int i; 1856 1857 for (i = 0; i < vi->curr_queue_pairs; i++) { 1858 struct receive_queue *rq = &vi->rq[i]; 1859 1860 napi_disable(&rq->napi); 1861 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL); 1862 virtnet_napi_enable(rq->vq, &rq->napi); 1863 1864 /* In theory, this can happen: if we don't get any buffers in 1865 * we will *never* try to fill again. 1866 */ 1867 if (still_empty) 1868 schedule_delayed_work(&vi->refill, HZ/2); 1869 } 1870 } 1871 1872 static int virtnet_receive(struct receive_queue *rq, int budget, 1873 unsigned int *xdp_xmit) 1874 { 1875 struct virtnet_info *vi = rq->vq->vdev->priv; 1876 struct virtnet_rq_stats stats = {}; 1877 unsigned int len; 1878 void *buf; 1879 int i; 1880 1881 if (!vi->big_packets || vi->mergeable_rx_bufs) { 1882 void *ctx; 1883 1884 while (stats.packets < budget && 1885 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) { 1886 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats); 1887 stats.packets++; 1888 } 1889 } else { 1890 while (stats.packets < budget && 1891 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) { 1892 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats); 1893 stats.packets++; 1894 } 1895 } 1896 1897 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) { 1898 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) { 1899 spin_lock(&vi->refill_lock); 1900 if (vi->refill_enabled) 1901 schedule_delayed_work(&vi->refill, 0); 1902 spin_unlock(&vi->refill_lock); 1903 } 1904 } 1905 1906 u64_stats_update_begin(&rq->stats.syncp); 1907 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) { 1908 size_t offset = virtnet_rq_stats_desc[i].offset; 1909 u64 *item; 1910 1911 item = (u64 *)((u8 *)&rq->stats + offset); 1912 *item += *(u64 *)((u8 *)&stats + offset); 1913 } 1914 u64_stats_update_end(&rq->stats.syncp); 1915 1916 return stats.packets; 1917 } 1918 1919 static void virtnet_poll_cleantx(struct receive_queue *rq) 1920 { 1921 struct virtnet_info *vi = rq->vq->vdev->priv; 1922 unsigned int index = vq2rxq(rq->vq); 1923 struct send_queue *sq = &vi->sq[index]; 1924 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index); 1925 1926 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index)) 1927 return; 1928 1929 if (__netif_tx_trylock(txq)) { 1930 if (sq->reset) { 1931 __netif_tx_unlock(txq); 1932 return; 1933 } 1934 1935 do { 1936 virtqueue_disable_cb(sq->vq); 1937 free_old_xmit_skbs(sq, true); 1938 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 1939 1940 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 1941 netif_tx_wake_queue(txq); 1942 1943 __netif_tx_unlock(txq); 1944 } 1945 } 1946 1947 static int virtnet_poll(struct napi_struct *napi, int budget) 1948 { 1949 struct receive_queue *rq = 1950 container_of(napi, struct receive_queue, napi); 1951 struct virtnet_info *vi = rq->vq->vdev->priv; 1952 struct send_queue *sq; 1953 unsigned int received; 1954 unsigned int xdp_xmit = 0; 1955 1956 virtnet_poll_cleantx(rq); 1957 1958 received = virtnet_receive(rq, budget, &xdp_xmit); 1959 1960 if (xdp_xmit & VIRTIO_XDP_REDIR) 1961 xdp_do_flush(); 1962 1963 /* Out of packets? */ 1964 if (received < budget) 1965 virtqueue_napi_complete(napi, rq->vq, received); 1966 1967 if (xdp_xmit & VIRTIO_XDP_TX) { 1968 sq = virtnet_xdp_get_sq(vi); 1969 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 1970 u64_stats_update_begin(&sq->stats.syncp); 1971 sq->stats.kicks++; 1972 u64_stats_update_end(&sq->stats.syncp); 1973 } 1974 virtnet_xdp_put_sq(vi, sq); 1975 } 1976 1977 return received; 1978 } 1979 1980 static int virtnet_open(struct net_device *dev) 1981 { 1982 struct virtnet_info *vi = netdev_priv(dev); 1983 int i, err; 1984 1985 enable_delayed_refill(vi); 1986 1987 for (i = 0; i < vi->max_queue_pairs; i++) { 1988 if (i < vi->curr_queue_pairs) 1989 /* Make sure we have some buffers: if oom use wq. */ 1990 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL)) 1991 schedule_delayed_work(&vi->refill, 0); 1992 1993 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id); 1994 if (err < 0) 1995 return err; 1996 1997 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq, 1998 MEM_TYPE_PAGE_SHARED, NULL); 1999 if (err < 0) { 2000 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 2001 return err; 2002 } 2003 2004 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 2005 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi); 2006 } 2007 2008 return 0; 2009 } 2010 2011 static int virtnet_poll_tx(struct napi_struct *napi, int budget) 2012 { 2013 struct send_queue *sq = container_of(napi, struct send_queue, napi); 2014 struct virtnet_info *vi = sq->vq->vdev->priv; 2015 unsigned int index = vq2txq(sq->vq); 2016 struct netdev_queue *txq; 2017 int opaque; 2018 bool done; 2019 2020 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) { 2021 /* We don't need to enable cb for XDP */ 2022 napi_complete_done(napi, 0); 2023 return 0; 2024 } 2025 2026 txq = netdev_get_tx_queue(vi->dev, index); 2027 __netif_tx_lock(txq, raw_smp_processor_id()); 2028 virtqueue_disable_cb(sq->vq); 2029 free_old_xmit_skbs(sq, true); 2030 2031 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) 2032 netif_tx_wake_queue(txq); 2033 2034 opaque = virtqueue_enable_cb_prepare(sq->vq); 2035 2036 done = napi_complete_done(napi, 0); 2037 2038 if (!done) 2039 virtqueue_disable_cb(sq->vq); 2040 2041 __netif_tx_unlock(txq); 2042 2043 if (done) { 2044 if (unlikely(virtqueue_poll(sq->vq, opaque))) { 2045 if (napi_schedule_prep(napi)) { 2046 __netif_tx_lock(txq, raw_smp_processor_id()); 2047 virtqueue_disable_cb(sq->vq); 2048 __netif_tx_unlock(txq); 2049 __napi_schedule(napi); 2050 } 2051 } 2052 } 2053 2054 return 0; 2055 } 2056 2057 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb) 2058 { 2059 struct virtio_net_hdr_mrg_rxbuf *hdr; 2060 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest; 2061 struct virtnet_info *vi = sq->vq->vdev->priv; 2062 int num_sg; 2063 unsigned hdr_len = vi->hdr_len; 2064 bool can_push; 2065 2066 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest); 2067 2068 can_push = vi->any_header_sg && 2069 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) && 2070 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len; 2071 /* Even if we can, don't push here yet as this would skew 2072 * csum_start offset below. */ 2073 if (can_push) 2074 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len); 2075 else 2076 hdr = skb_vnet_hdr(skb); 2077 2078 if (virtio_net_hdr_from_skb(skb, &hdr->hdr, 2079 virtio_is_little_endian(vi->vdev), false, 2080 0)) 2081 return -EPROTO; 2082 2083 if (vi->mergeable_rx_bufs) 2084 hdr->num_buffers = 0; 2085 2086 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2)); 2087 if (can_push) { 2088 __skb_push(skb, hdr_len); 2089 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len); 2090 if (unlikely(num_sg < 0)) 2091 return num_sg; 2092 /* Pull header back to avoid skew in tx bytes calculations. */ 2093 __skb_pull(skb, hdr_len); 2094 } else { 2095 sg_set_buf(sq->sg, hdr, hdr_len); 2096 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len); 2097 if (unlikely(num_sg < 0)) 2098 return num_sg; 2099 num_sg++; 2100 } 2101 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC); 2102 } 2103 2104 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev) 2105 { 2106 struct virtnet_info *vi = netdev_priv(dev); 2107 int qnum = skb_get_queue_mapping(skb); 2108 struct send_queue *sq = &vi->sq[qnum]; 2109 int err; 2110 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum); 2111 bool kick = !netdev_xmit_more(); 2112 bool use_napi = sq->napi.weight; 2113 2114 /* Free up any pending old buffers before queueing new ones. */ 2115 do { 2116 if (use_napi) 2117 virtqueue_disable_cb(sq->vq); 2118 2119 free_old_xmit_skbs(sq, false); 2120 2121 } while (use_napi && kick && 2122 unlikely(!virtqueue_enable_cb_delayed(sq->vq))); 2123 2124 /* timestamp packet in software */ 2125 skb_tx_timestamp(skb); 2126 2127 /* Try to transmit */ 2128 err = xmit_skb(sq, skb); 2129 2130 /* This should not happen! */ 2131 if (unlikely(err)) { 2132 dev->stats.tx_fifo_errors++; 2133 if (net_ratelimit()) 2134 dev_warn(&dev->dev, 2135 "Unexpected TXQ (%d) queue failure: %d\n", 2136 qnum, err); 2137 dev->stats.tx_dropped++; 2138 dev_kfree_skb_any(skb); 2139 return NETDEV_TX_OK; 2140 } 2141 2142 /* Don't wait up for transmitted skbs to be freed. */ 2143 if (!use_napi) { 2144 skb_orphan(skb); 2145 nf_reset_ct(skb); 2146 } 2147 2148 check_sq_full_and_disable(vi, dev, sq); 2149 2150 if (kick || netif_xmit_stopped(txq)) { 2151 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) { 2152 u64_stats_update_begin(&sq->stats.syncp); 2153 sq->stats.kicks++; 2154 u64_stats_update_end(&sq->stats.syncp); 2155 } 2156 } 2157 2158 return NETDEV_TX_OK; 2159 } 2160 2161 static int virtnet_rx_resize(struct virtnet_info *vi, 2162 struct receive_queue *rq, u32 ring_num) 2163 { 2164 bool running = netif_running(vi->dev); 2165 int err, qindex; 2166 2167 qindex = rq - vi->rq; 2168 2169 if (running) 2170 napi_disable(&rq->napi); 2171 2172 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf); 2173 if (err) 2174 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err); 2175 2176 if (!try_fill_recv(vi, rq, GFP_KERNEL)) 2177 schedule_delayed_work(&vi->refill, 0); 2178 2179 if (running) 2180 virtnet_napi_enable(rq->vq, &rq->napi); 2181 return err; 2182 } 2183 2184 static int virtnet_tx_resize(struct virtnet_info *vi, 2185 struct send_queue *sq, u32 ring_num) 2186 { 2187 bool running = netif_running(vi->dev); 2188 struct netdev_queue *txq; 2189 int err, qindex; 2190 2191 qindex = sq - vi->sq; 2192 2193 if (running) 2194 virtnet_napi_tx_disable(&sq->napi); 2195 2196 txq = netdev_get_tx_queue(vi->dev, qindex); 2197 2198 /* 1. wait all ximt complete 2199 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue() 2200 */ 2201 __netif_tx_lock_bh(txq); 2202 2203 /* Prevent rx poll from accessing sq. */ 2204 sq->reset = true; 2205 2206 /* Prevent the upper layer from trying to send packets. */ 2207 netif_stop_subqueue(vi->dev, qindex); 2208 2209 __netif_tx_unlock_bh(txq); 2210 2211 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf); 2212 if (err) 2213 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err); 2214 2215 __netif_tx_lock_bh(txq); 2216 sq->reset = false; 2217 netif_tx_wake_queue(txq); 2218 __netif_tx_unlock_bh(txq); 2219 2220 if (running) 2221 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi); 2222 return err; 2223 } 2224 2225 /* 2226 * Send command via the control virtqueue and check status. Commands 2227 * supported by the hypervisor, as indicated by feature bits, should 2228 * never fail unless improperly formatted. 2229 */ 2230 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd, 2231 struct scatterlist *out) 2232 { 2233 struct scatterlist *sgs[4], hdr, stat; 2234 unsigned out_num = 0, tmp; 2235 int ret; 2236 2237 /* Caller should know better */ 2238 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ)); 2239 2240 vi->ctrl->status = ~0; 2241 vi->ctrl->hdr.class = class; 2242 vi->ctrl->hdr.cmd = cmd; 2243 /* Add header */ 2244 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr)); 2245 sgs[out_num++] = &hdr; 2246 2247 if (out) 2248 sgs[out_num++] = out; 2249 2250 /* Add return status. */ 2251 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status)); 2252 sgs[out_num] = &stat; 2253 2254 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs)); 2255 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC); 2256 if (ret < 0) { 2257 dev_warn(&vi->vdev->dev, 2258 "Failed to add sgs for command vq: %d\n.", ret); 2259 return false; 2260 } 2261 2262 if (unlikely(!virtqueue_kick(vi->cvq))) 2263 return vi->ctrl->status == VIRTIO_NET_OK; 2264 2265 /* Spin for a response, the kick causes an ioport write, trapping 2266 * into the hypervisor, so the request should be handled immediately. 2267 */ 2268 while (!virtqueue_get_buf(vi->cvq, &tmp) && 2269 !virtqueue_is_broken(vi->cvq)) 2270 cpu_relax(); 2271 2272 return vi->ctrl->status == VIRTIO_NET_OK; 2273 } 2274 2275 static int virtnet_set_mac_address(struct net_device *dev, void *p) 2276 { 2277 struct virtnet_info *vi = netdev_priv(dev); 2278 struct virtio_device *vdev = vi->vdev; 2279 int ret; 2280 struct sockaddr *addr; 2281 struct scatterlist sg; 2282 2283 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 2284 return -EOPNOTSUPP; 2285 2286 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL); 2287 if (!addr) 2288 return -ENOMEM; 2289 2290 ret = eth_prepare_mac_addr_change(dev, addr); 2291 if (ret) 2292 goto out; 2293 2294 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 2295 sg_init_one(&sg, addr->sa_data, dev->addr_len); 2296 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 2297 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 2298 dev_warn(&vdev->dev, 2299 "Failed to set mac address by vq command.\n"); 2300 ret = -EINVAL; 2301 goto out; 2302 } 2303 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 2304 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) { 2305 unsigned int i; 2306 2307 /* Naturally, this has an atomicity problem. */ 2308 for (i = 0; i < dev->addr_len; i++) 2309 virtio_cwrite8(vdev, 2310 offsetof(struct virtio_net_config, mac) + 2311 i, addr->sa_data[i]); 2312 } 2313 2314 eth_commit_mac_addr_change(dev, p); 2315 ret = 0; 2316 2317 out: 2318 kfree(addr); 2319 return ret; 2320 } 2321 2322 static void virtnet_stats(struct net_device *dev, 2323 struct rtnl_link_stats64 *tot) 2324 { 2325 struct virtnet_info *vi = netdev_priv(dev); 2326 unsigned int start; 2327 int i; 2328 2329 for (i = 0; i < vi->max_queue_pairs; i++) { 2330 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops; 2331 struct receive_queue *rq = &vi->rq[i]; 2332 struct send_queue *sq = &vi->sq[i]; 2333 2334 do { 2335 start = u64_stats_fetch_begin(&sq->stats.syncp); 2336 tpackets = sq->stats.packets; 2337 tbytes = sq->stats.bytes; 2338 terrors = sq->stats.tx_timeouts; 2339 } while (u64_stats_fetch_retry(&sq->stats.syncp, start)); 2340 2341 do { 2342 start = u64_stats_fetch_begin(&rq->stats.syncp); 2343 rpackets = rq->stats.packets; 2344 rbytes = rq->stats.bytes; 2345 rdrops = rq->stats.drops; 2346 } while (u64_stats_fetch_retry(&rq->stats.syncp, start)); 2347 2348 tot->rx_packets += rpackets; 2349 tot->tx_packets += tpackets; 2350 tot->rx_bytes += rbytes; 2351 tot->tx_bytes += tbytes; 2352 tot->rx_dropped += rdrops; 2353 tot->tx_errors += terrors; 2354 } 2355 2356 tot->tx_dropped = dev->stats.tx_dropped; 2357 tot->tx_fifo_errors = dev->stats.tx_fifo_errors; 2358 tot->rx_length_errors = dev->stats.rx_length_errors; 2359 tot->rx_frame_errors = dev->stats.rx_frame_errors; 2360 } 2361 2362 static void virtnet_ack_link_announce(struct virtnet_info *vi) 2363 { 2364 rtnl_lock(); 2365 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE, 2366 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL)) 2367 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n"); 2368 rtnl_unlock(); 2369 } 2370 2371 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2372 { 2373 struct scatterlist sg; 2374 struct net_device *dev = vi->dev; 2375 2376 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ)) 2377 return 0; 2378 2379 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs); 2380 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq)); 2381 2382 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2383 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) { 2384 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n", 2385 queue_pairs); 2386 return -EINVAL; 2387 } else { 2388 vi->curr_queue_pairs = queue_pairs; 2389 /* virtnet_open() will refill when device is going to up. */ 2390 if (dev->flags & IFF_UP) 2391 schedule_delayed_work(&vi->refill, 0); 2392 } 2393 2394 return 0; 2395 } 2396 2397 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs) 2398 { 2399 int err; 2400 2401 rtnl_lock(); 2402 err = _virtnet_set_queues(vi, queue_pairs); 2403 rtnl_unlock(); 2404 return err; 2405 } 2406 2407 static int virtnet_close(struct net_device *dev) 2408 { 2409 struct virtnet_info *vi = netdev_priv(dev); 2410 int i; 2411 2412 /* Make sure NAPI doesn't schedule refill work */ 2413 disable_delayed_refill(vi); 2414 /* Make sure refill_work doesn't re-enable napi! */ 2415 cancel_delayed_work_sync(&vi->refill); 2416 2417 for (i = 0; i < vi->max_queue_pairs; i++) { 2418 virtnet_napi_tx_disable(&vi->sq[i].napi); 2419 napi_disable(&vi->rq[i].napi); 2420 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq); 2421 } 2422 2423 return 0; 2424 } 2425 2426 static void virtnet_set_rx_mode(struct net_device *dev) 2427 { 2428 struct virtnet_info *vi = netdev_priv(dev); 2429 struct scatterlist sg[2]; 2430 struct virtio_net_ctrl_mac *mac_data; 2431 struct netdev_hw_addr *ha; 2432 int uc_count; 2433 int mc_count; 2434 void *buf; 2435 int i; 2436 2437 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */ 2438 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX)) 2439 return; 2440 2441 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0); 2442 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0); 2443 2444 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc)); 2445 2446 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2447 VIRTIO_NET_CTRL_RX_PROMISC, sg)) 2448 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n", 2449 vi->ctrl->promisc ? "en" : "dis"); 2450 2451 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti)); 2452 2453 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX, 2454 VIRTIO_NET_CTRL_RX_ALLMULTI, sg)) 2455 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n", 2456 vi->ctrl->allmulti ? "en" : "dis"); 2457 2458 uc_count = netdev_uc_count(dev); 2459 mc_count = netdev_mc_count(dev); 2460 /* MAC filter - use one buffer for both lists */ 2461 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) + 2462 (2 * sizeof(mac_data->entries)), GFP_ATOMIC); 2463 mac_data = buf; 2464 if (!buf) 2465 return; 2466 2467 sg_init_table(sg, 2); 2468 2469 /* Store the unicast list and count in the front of the buffer */ 2470 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count); 2471 i = 0; 2472 netdev_for_each_uc_addr(ha, dev) 2473 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2474 2475 sg_set_buf(&sg[0], mac_data, 2476 sizeof(mac_data->entries) + (uc_count * ETH_ALEN)); 2477 2478 /* multicast list and count fill the end */ 2479 mac_data = (void *)&mac_data->macs[uc_count][0]; 2480 2481 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count); 2482 i = 0; 2483 netdev_for_each_mc_addr(ha, dev) 2484 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN); 2485 2486 sg_set_buf(&sg[1], mac_data, 2487 sizeof(mac_data->entries) + (mc_count * ETH_ALEN)); 2488 2489 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 2490 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg)) 2491 dev_warn(&dev->dev, "Failed to set MAC filter table.\n"); 2492 2493 kfree(buf); 2494 } 2495 2496 static int virtnet_vlan_rx_add_vid(struct net_device *dev, 2497 __be16 proto, u16 vid) 2498 { 2499 struct virtnet_info *vi = netdev_priv(dev); 2500 struct scatterlist sg; 2501 2502 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2503 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2504 2505 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2506 VIRTIO_NET_CTRL_VLAN_ADD, &sg)) 2507 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid); 2508 return 0; 2509 } 2510 2511 static int virtnet_vlan_rx_kill_vid(struct net_device *dev, 2512 __be16 proto, u16 vid) 2513 { 2514 struct virtnet_info *vi = netdev_priv(dev); 2515 struct scatterlist sg; 2516 2517 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid); 2518 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid)); 2519 2520 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN, 2521 VIRTIO_NET_CTRL_VLAN_DEL, &sg)) 2522 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid); 2523 return 0; 2524 } 2525 2526 static void virtnet_clean_affinity(struct virtnet_info *vi) 2527 { 2528 int i; 2529 2530 if (vi->affinity_hint_set) { 2531 for (i = 0; i < vi->max_queue_pairs; i++) { 2532 virtqueue_set_affinity(vi->rq[i].vq, NULL); 2533 virtqueue_set_affinity(vi->sq[i].vq, NULL); 2534 } 2535 2536 vi->affinity_hint_set = false; 2537 } 2538 } 2539 2540 static void virtnet_set_affinity(struct virtnet_info *vi) 2541 { 2542 cpumask_var_t mask; 2543 int stragglers; 2544 int group_size; 2545 int i, j, cpu; 2546 int num_cpu; 2547 int stride; 2548 2549 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) { 2550 virtnet_clean_affinity(vi); 2551 return; 2552 } 2553 2554 num_cpu = num_online_cpus(); 2555 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1); 2556 stragglers = num_cpu >= vi->curr_queue_pairs ? 2557 num_cpu % vi->curr_queue_pairs : 2558 0; 2559 cpu = cpumask_first(cpu_online_mask); 2560 2561 for (i = 0; i < vi->curr_queue_pairs; i++) { 2562 group_size = stride + (i < stragglers ? 1 : 0); 2563 2564 for (j = 0; j < group_size; j++) { 2565 cpumask_set_cpu(cpu, mask); 2566 cpu = cpumask_next_wrap(cpu, cpu_online_mask, 2567 nr_cpu_ids, false); 2568 } 2569 virtqueue_set_affinity(vi->rq[i].vq, mask); 2570 virtqueue_set_affinity(vi->sq[i].vq, mask); 2571 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS); 2572 cpumask_clear(mask); 2573 } 2574 2575 vi->affinity_hint_set = true; 2576 free_cpumask_var(mask); 2577 } 2578 2579 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node) 2580 { 2581 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2582 node); 2583 virtnet_set_affinity(vi); 2584 return 0; 2585 } 2586 2587 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node) 2588 { 2589 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2590 node_dead); 2591 virtnet_set_affinity(vi); 2592 return 0; 2593 } 2594 2595 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node) 2596 { 2597 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info, 2598 node); 2599 2600 virtnet_clean_affinity(vi); 2601 return 0; 2602 } 2603 2604 static enum cpuhp_state virtionet_online; 2605 2606 static int virtnet_cpu_notif_add(struct virtnet_info *vi) 2607 { 2608 int ret; 2609 2610 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node); 2611 if (ret) 2612 return ret; 2613 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2614 &vi->node_dead); 2615 if (!ret) 2616 return ret; 2617 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2618 return ret; 2619 } 2620 2621 static void virtnet_cpu_notif_remove(struct virtnet_info *vi) 2622 { 2623 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node); 2624 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD, 2625 &vi->node_dead); 2626 } 2627 2628 static void virtnet_get_ringparam(struct net_device *dev, 2629 struct ethtool_ringparam *ring, 2630 struct kernel_ethtool_ringparam *kernel_ring, 2631 struct netlink_ext_ack *extack) 2632 { 2633 struct virtnet_info *vi = netdev_priv(dev); 2634 2635 ring->rx_max_pending = vi->rq[0].vq->num_max; 2636 ring->tx_max_pending = vi->sq[0].vq->num_max; 2637 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq); 2638 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq); 2639 } 2640 2641 static int virtnet_set_ringparam(struct net_device *dev, 2642 struct ethtool_ringparam *ring, 2643 struct kernel_ethtool_ringparam *kernel_ring, 2644 struct netlink_ext_ack *extack) 2645 { 2646 struct virtnet_info *vi = netdev_priv(dev); 2647 u32 rx_pending, tx_pending; 2648 struct receive_queue *rq; 2649 struct send_queue *sq; 2650 int i, err; 2651 2652 if (ring->rx_mini_pending || ring->rx_jumbo_pending) 2653 return -EINVAL; 2654 2655 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq); 2656 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq); 2657 2658 if (ring->rx_pending == rx_pending && 2659 ring->tx_pending == tx_pending) 2660 return 0; 2661 2662 if (ring->rx_pending > vi->rq[0].vq->num_max) 2663 return -EINVAL; 2664 2665 if (ring->tx_pending > vi->sq[0].vq->num_max) 2666 return -EINVAL; 2667 2668 for (i = 0; i < vi->max_queue_pairs; i++) { 2669 rq = vi->rq + i; 2670 sq = vi->sq + i; 2671 2672 if (ring->tx_pending != tx_pending) { 2673 err = virtnet_tx_resize(vi, sq, ring->tx_pending); 2674 if (err) 2675 return err; 2676 } 2677 2678 if (ring->rx_pending != rx_pending) { 2679 err = virtnet_rx_resize(vi, rq, ring->rx_pending); 2680 if (err) 2681 return err; 2682 } 2683 } 2684 2685 return 0; 2686 } 2687 2688 static bool virtnet_commit_rss_command(struct virtnet_info *vi) 2689 { 2690 struct net_device *dev = vi->dev; 2691 struct scatterlist sgs[4]; 2692 unsigned int sg_buf_size; 2693 2694 /* prepare sgs */ 2695 sg_init_table(sgs, 4); 2696 2697 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table); 2698 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size); 2699 2700 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1); 2701 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size); 2702 2703 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key) 2704 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq); 2705 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size); 2706 2707 sg_buf_size = vi->rss_key_size; 2708 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size); 2709 2710 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ, 2711 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG 2712 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) { 2713 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n"); 2714 return false; 2715 } 2716 return true; 2717 } 2718 2719 static void virtnet_init_default_rss(struct virtnet_info *vi) 2720 { 2721 u32 indir_val = 0; 2722 int i = 0; 2723 2724 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported; 2725 vi->rss_hash_types_saved = vi->rss_hash_types_supported; 2726 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size 2727 ? vi->rss_indir_table_size - 1 : 0; 2728 vi->ctrl->rss.unclassified_queue = 0; 2729 2730 for (; i < vi->rss_indir_table_size; ++i) { 2731 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs); 2732 vi->ctrl->rss.indirection_table[i] = indir_val; 2733 } 2734 2735 vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs; 2736 vi->ctrl->rss.hash_key_length = vi->rss_key_size; 2737 2738 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size); 2739 } 2740 2741 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info) 2742 { 2743 info->data = 0; 2744 switch (info->flow_type) { 2745 case TCP_V4_FLOW: 2746 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) { 2747 info->data = RXH_IP_SRC | RXH_IP_DST | 2748 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2749 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2750 info->data = RXH_IP_SRC | RXH_IP_DST; 2751 } 2752 break; 2753 case TCP_V6_FLOW: 2754 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) { 2755 info->data = RXH_IP_SRC | RXH_IP_DST | 2756 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2757 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2758 info->data = RXH_IP_SRC | RXH_IP_DST; 2759 } 2760 break; 2761 case UDP_V4_FLOW: 2762 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) { 2763 info->data = RXH_IP_SRC | RXH_IP_DST | 2764 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2765 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) { 2766 info->data = RXH_IP_SRC | RXH_IP_DST; 2767 } 2768 break; 2769 case UDP_V6_FLOW: 2770 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) { 2771 info->data = RXH_IP_SRC | RXH_IP_DST | 2772 RXH_L4_B_0_1 | RXH_L4_B_2_3; 2773 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) { 2774 info->data = RXH_IP_SRC | RXH_IP_DST; 2775 } 2776 break; 2777 case IPV4_FLOW: 2778 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) 2779 info->data = RXH_IP_SRC | RXH_IP_DST; 2780 2781 break; 2782 case IPV6_FLOW: 2783 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) 2784 info->data = RXH_IP_SRC | RXH_IP_DST; 2785 2786 break; 2787 default: 2788 info->data = 0; 2789 break; 2790 } 2791 } 2792 2793 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info) 2794 { 2795 u32 new_hashtypes = vi->rss_hash_types_saved; 2796 bool is_disable = info->data & RXH_DISCARD; 2797 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3); 2798 2799 /* supports only 'sd', 'sdfn' and 'r' */ 2800 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable)) 2801 return false; 2802 2803 switch (info->flow_type) { 2804 case TCP_V4_FLOW: 2805 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4); 2806 if (!is_disable) 2807 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2808 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0); 2809 break; 2810 case UDP_V4_FLOW: 2811 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4); 2812 if (!is_disable) 2813 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4 2814 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0); 2815 break; 2816 case IPV4_FLOW: 2817 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2818 if (!is_disable) 2819 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4; 2820 break; 2821 case TCP_V6_FLOW: 2822 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6); 2823 if (!is_disable) 2824 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2825 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0); 2826 break; 2827 case UDP_V6_FLOW: 2828 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6); 2829 if (!is_disable) 2830 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6 2831 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0); 2832 break; 2833 case IPV6_FLOW: 2834 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2835 if (!is_disable) 2836 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6; 2837 break; 2838 default: 2839 /* unsupported flow */ 2840 return false; 2841 } 2842 2843 /* if unsupported hashtype was set */ 2844 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported)) 2845 return false; 2846 2847 if (new_hashtypes != vi->rss_hash_types_saved) { 2848 vi->rss_hash_types_saved = new_hashtypes; 2849 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 2850 if (vi->dev->features & NETIF_F_RXHASH) 2851 return virtnet_commit_rss_command(vi); 2852 } 2853 2854 return true; 2855 } 2856 2857 static void virtnet_get_drvinfo(struct net_device *dev, 2858 struct ethtool_drvinfo *info) 2859 { 2860 struct virtnet_info *vi = netdev_priv(dev); 2861 struct virtio_device *vdev = vi->vdev; 2862 2863 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver)); 2864 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version)); 2865 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info)); 2866 2867 } 2868 2869 /* TODO: Eliminate OOO packets during switching */ 2870 static int virtnet_set_channels(struct net_device *dev, 2871 struct ethtool_channels *channels) 2872 { 2873 struct virtnet_info *vi = netdev_priv(dev); 2874 u16 queue_pairs = channels->combined_count; 2875 int err; 2876 2877 /* We don't support separate rx/tx channels. 2878 * We don't allow setting 'other' channels. 2879 */ 2880 if (channels->rx_count || channels->tx_count || channels->other_count) 2881 return -EINVAL; 2882 2883 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0) 2884 return -EINVAL; 2885 2886 /* For now we don't support modifying channels while XDP is loaded 2887 * also when XDP is loaded all RX queues have XDP programs so we only 2888 * need to check a single RX queue. 2889 */ 2890 if (vi->rq[0].xdp_prog) 2891 return -EINVAL; 2892 2893 cpus_read_lock(); 2894 err = _virtnet_set_queues(vi, queue_pairs); 2895 if (err) { 2896 cpus_read_unlock(); 2897 goto err; 2898 } 2899 virtnet_set_affinity(vi); 2900 cpus_read_unlock(); 2901 2902 netif_set_real_num_tx_queues(dev, queue_pairs); 2903 netif_set_real_num_rx_queues(dev, queue_pairs); 2904 err: 2905 return err; 2906 } 2907 2908 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data) 2909 { 2910 struct virtnet_info *vi = netdev_priv(dev); 2911 unsigned int i, j; 2912 u8 *p = data; 2913 2914 switch (stringset) { 2915 case ETH_SS_STATS: 2916 for (i = 0; i < vi->curr_queue_pairs; i++) { 2917 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) 2918 ethtool_sprintf(&p, "rx_queue_%u_%s", i, 2919 virtnet_rq_stats_desc[j].desc); 2920 } 2921 2922 for (i = 0; i < vi->curr_queue_pairs; i++) { 2923 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) 2924 ethtool_sprintf(&p, "tx_queue_%u_%s", i, 2925 virtnet_sq_stats_desc[j].desc); 2926 } 2927 break; 2928 } 2929 } 2930 2931 static int virtnet_get_sset_count(struct net_device *dev, int sset) 2932 { 2933 struct virtnet_info *vi = netdev_priv(dev); 2934 2935 switch (sset) { 2936 case ETH_SS_STATS: 2937 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN + 2938 VIRTNET_SQ_STATS_LEN); 2939 default: 2940 return -EOPNOTSUPP; 2941 } 2942 } 2943 2944 static void virtnet_get_ethtool_stats(struct net_device *dev, 2945 struct ethtool_stats *stats, u64 *data) 2946 { 2947 struct virtnet_info *vi = netdev_priv(dev); 2948 unsigned int idx = 0, start, i, j; 2949 const u8 *stats_base; 2950 size_t offset; 2951 2952 for (i = 0; i < vi->curr_queue_pairs; i++) { 2953 struct receive_queue *rq = &vi->rq[i]; 2954 2955 stats_base = (u8 *)&rq->stats; 2956 do { 2957 start = u64_stats_fetch_begin(&rq->stats.syncp); 2958 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) { 2959 offset = virtnet_rq_stats_desc[j].offset; 2960 data[idx + j] = *(u64 *)(stats_base + offset); 2961 } 2962 } while (u64_stats_fetch_retry(&rq->stats.syncp, start)); 2963 idx += VIRTNET_RQ_STATS_LEN; 2964 } 2965 2966 for (i = 0; i < vi->curr_queue_pairs; i++) { 2967 struct send_queue *sq = &vi->sq[i]; 2968 2969 stats_base = (u8 *)&sq->stats; 2970 do { 2971 start = u64_stats_fetch_begin(&sq->stats.syncp); 2972 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) { 2973 offset = virtnet_sq_stats_desc[j].offset; 2974 data[idx + j] = *(u64 *)(stats_base + offset); 2975 } 2976 } while (u64_stats_fetch_retry(&sq->stats.syncp, start)); 2977 idx += VIRTNET_SQ_STATS_LEN; 2978 } 2979 } 2980 2981 static void virtnet_get_channels(struct net_device *dev, 2982 struct ethtool_channels *channels) 2983 { 2984 struct virtnet_info *vi = netdev_priv(dev); 2985 2986 channels->combined_count = vi->curr_queue_pairs; 2987 channels->max_combined = vi->max_queue_pairs; 2988 channels->max_other = 0; 2989 channels->rx_count = 0; 2990 channels->tx_count = 0; 2991 channels->other_count = 0; 2992 } 2993 2994 static int virtnet_set_link_ksettings(struct net_device *dev, 2995 const struct ethtool_link_ksettings *cmd) 2996 { 2997 struct virtnet_info *vi = netdev_priv(dev); 2998 2999 return ethtool_virtdev_set_link_ksettings(dev, cmd, 3000 &vi->speed, &vi->duplex); 3001 } 3002 3003 static int virtnet_get_link_ksettings(struct net_device *dev, 3004 struct ethtool_link_ksettings *cmd) 3005 { 3006 struct virtnet_info *vi = netdev_priv(dev); 3007 3008 cmd->base.speed = vi->speed; 3009 cmd->base.duplex = vi->duplex; 3010 cmd->base.port = PORT_OTHER; 3011 3012 return 0; 3013 } 3014 3015 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi, 3016 struct ethtool_coalesce *ec) 3017 { 3018 struct scatterlist sgs_tx, sgs_rx; 3019 struct virtio_net_ctrl_coal_tx coal_tx; 3020 struct virtio_net_ctrl_coal_rx coal_rx; 3021 3022 coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs); 3023 coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames); 3024 sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx)); 3025 3026 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL, 3027 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET, 3028 &sgs_tx)) 3029 return -EINVAL; 3030 3031 /* Save parameters */ 3032 vi->tx_usecs = ec->tx_coalesce_usecs; 3033 vi->tx_max_packets = ec->tx_max_coalesced_frames; 3034 3035 coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs); 3036 coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames); 3037 sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx)); 3038 3039 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL, 3040 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET, 3041 &sgs_rx)) 3042 return -EINVAL; 3043 3044 /* Save parameters */ 3045 vi->rx_usecs = ec->rx_coalesce_usecs; 3046 vi->rx_max_packets = ec->rx_max_coalesced_frames; 3047 3048 return 0; 3049 } 3050 3051 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec) 3052 { 3053 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL 3054 * feature is negotiated. 3055 */ 3056 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs) 3057 return -EOPNOTSUPP; 3058 3059 if (ec->tx_max_coalesced_frames > 1 || 3060 ec->rx_max_coalesced_frames != 1) 3061 return -EINVAL; 3062 3063 return 0; 3064 } 3065 3066 static int virtnet_set_coalesce(struct net_device *dev, 3067 struct ethtool_coalesce *ec, 3068 struct kernel_ethtool_coalesce *kernel_coal, 3069 struct netlink_ext_ack *extack) 3070 { 3071 struct virtnet_info *vi = netdev_priv(dev); 3072 int ret, i, napi_weight; 3073 bool update_napi = false; 3074 3075 /* Can't change NAPI weight if the link is up */ 3076 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0; 3077 if (napi_weight ^ vi->sq[0].napi.weight) { 3078 if (dev->flags & IFF_UP) 3079 return -EBUSY; 3080 else 3081 update_napi = true; 3082 } 3083 3084 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) 3085 ret = virtnet_send_notf_coal_cmds(vi, ec); 3086 else 3087 ret = virtnet_coal_params_supported(ec); 3088 3089 if (ret) 3090 return ret; 3091 3092 if (update_napi) { 3093 for (i = 0; i < vi->max_queue_pairs; i++) 3094 vi->sq[i].napi.weight = napi_weight; 3095 } 3096 3097 return ret; 3098 } 3099 3100 static int virtnet_get_coalesce(struct net_device *dev, 3101 struct ethtool_coalesce *ec, 3102 struct kernel_ethtool_coalesce *kernel_coal, 3103 struct netlink_ext_ack *extack) 3104 { 3105 struct virtnet_info *vi = netdev_priv(dev); 3106 3107 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) { 3108 ec->rx_coalesce_usecs = vi->rx_usecs; 3109 ec->tx_coalesce_usecs = vi->tx_usecs; 3110 ec->tx_max_coalesced_frames = vi->tx_max_packets; 3111 ec->rx_max_coalesced_frames = vi->rx_max_packets; 3112 } else { 3113 ec->rx_max_coalesced_frames = 1; 3114 3115 if (vi->sq[0].napi.weight) 3116 ec->tx_max_coalesced_frames = 1; 3117 } 3118 3119 return 0; 3120 } 3121 3122 static void virtnet_init_settings(struct net_device *dev) 3123 { 3124 struct virtnet_info *vi = netdev_priv(dev); 3125 3126 vi->speed = SPEED_UNKNOWN; 3127 vi->duplex = DUPLEX_UNKNOWN; 3128 } 3129 3130 static void virtnet_update_settings(struct virtnet_info *vi) 3131 { 3132 u32 speed; 3133 u8 duplex; 3134 3135 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX)) 3136 return; 3137 3138 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed); 3139 3140 if (ethtool_validate_speed(speed)) 3141 vi->speed = speed; 3142 3143 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex); 3144 3145 if (ethtool_validate_duplex(duplex)) 3146 vi->duplex = duplex; 3147 } 3148 3149 static u32 virtnet_get_rxfh_key_size(struct net_device *dev) 3150 { 3151 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size; 3152 } 3153 3154 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev) 3155 { 3156 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size; 3157 } 3158 3159 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc) 3160 { 3161 struct virtnet_info *vi = netdev_priv(dev); 3162 int i; 3163 3164 if (indir) { 3165 for (i = 0; i < vi->rss_indir_table_size; ++i) 3166 indir[i] = vi->ctrl->rss.indirection_table[i]; 3167 } 3168 3169 if (key) 3170 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size); 3171 3172 if (hfunc) 3173 *hfunc = ETH_RSS_HASH_TOP; 3174 3175 return 0; 3176 } 3177 3178 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc) 3179 { 3180 struct virtnet_info *vi = netdev_priv(dev); 3181 int i; 3182 3183 if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP) 3184 return -EOPNOTSUPP; 3185 3186 if (indir) { 3187 for (i = 0; i < vi->rss_indir_table_size; ++i) 3188 vi->ctrl->rss.indirection_table[i] = indir[i]; 3189 } 3190 if (key) 3191 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size); 3192 3193 virtnet_commit_rss_command(vi); 3194 3195 return 0; 3196 } 3197 3198 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs) 3199 { 3200 struct virtnet_info *vi = netdev_priv(dev); 3201 int rc = 0; 3202 3203 switch (info->cmd) { 3204 case ETHTOOL_GRXRINGS: 3205 info->data = vi->curr_queue_pairs; 3206 break; 3207 case ETHTOOL_GRXFH: 3208 virtnet_get_hashflow(vi, info); 3209 break; 3210 default: 3211 rc = -EOPNOTSUPP; 3212 } 3213 3214 return rc; 3215 } 3216 3217 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info) 3218 { 3219 struct virtnet_info *vi = netdev_priv(dev); 3220 int rc = 0; 3221 3222 switch (info->cmd) { 3223 case ETHTOOL_SRXFH: 3224 if (!virtnet_set_hashflow(vi, info)) 3225 rc = -EINVAL; 3226 3227 break; 3228 default: 3229 rc = -EOPNOTSUPP; 3230 } 3231 3232 return rc; 3233 } 3234 3235 static const struct ethtool_ops virtnet_ethtool_ops = { 3236 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES | 3237 ETHTOOL_COALESCE_USECS, 3238 .get_drvinfo = virtnet_get_drvinfo, 3239 .get_link = ethtool_op_get_link, 3240 .get_ringparam = virtnet_get_ringparam, 3241 .set_ringparam = virtnet_set_ringparam, 3242 .get_strings = virtnet_get_strings, 3243 .get_sset_count = virtnet_get_sset_count, 3244 .get_ethtool_stats = virtnet_get_ethtool_stats, 3245 .set_channels = virtnet_set_channels, 3246 .get_channels = virtnet_get_channels, 3247 .get_ts_info = ethtool_op_get_ts_info, 3248 .get_link_ksettings = virtnet_get_link_ksettings, 3249 .set_link_ksettings = virtnet_set_link_ksettings, 3250 .set_coalesce = virtnet_set_coalesce, 3251 .get_coalesce = virtnet_get_coalesce, 3252 .get_rxfh_key_size = virtnet_get_rxfh_key_size, 3253 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size, 3254 .get_rxfh = virtnet_get_rxfh, 3255 .set_rxfh = virtnet_set_rxfh, 3256 .get_rxnfc = virtnet_get_rxnfc, 3257 .set_rxnfc = virtnet_set_rxnfc, 3258 }; 3259 3260 static void virtnet_freeze_down(struct virtio_device *vdev) 3261 { 3262 struct virtnet_info *vi = vdev->priv; 3263 3264 /* Make sure no work handler is accessing the device */ 3265 flush_work(&vi->config_work); 3266 3267 netif_tx_lock_bh(vi->dev); 3268 netif_device_detach(vi->dev); 3269 netif_tx_unlock_bh(vi->dev); 3270 if (netif_running(vi->dev)) 3271 virtnet_close(vi->dev); 3272 } 3273 3274 static int init_vqs(struct virtnet_info *vi); 3275 3276 static int virtnet_restore_up(struct virtio_device *vdev) 3277 { 3278 struct virtnet_info *vi = vdev->priv; 3279 int err; 3280 3281 err = init_vqs(vi); 3282 if (err) 3283 return err; 3284 3285 virtio_device_ready(vdev); 3286 3287 enable_delayed_refill(vi); 3288 3289 if (netif_running(vi->dev)) { 3290 err = virtnet_open(vi->dev); 3291 if (err) 3292 return err; 3293 } 3294 3295 netif_tx_lock_bh(vi->dev); 3296 netif_device_attach(vi->dev); 3297 netif_tx_unlock_bh(vi->dev); 3298 return err; 3299 } 3300 3301 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads) 3302 { 3303 struct scatterlist sg; 3304 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads); 3305 3306 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads)); 3307 3308 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 3309 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) { 3310 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n"); 3311 return -EINVAL; 3312 } 3313 3314 return 0; 3315 } 3316 3317 static int virtnet_clear_guest_offloads(struct virtnet_info *vi) 3318 { 3319 u64 offloads = 0; 3320 3321 if (!vi->guest_offloads) 3322 return 0; 3323 3324 return virtnet_set_guest_offloads(vi, offloads); 3325 } 3326 3327 static int virtnet_restore_guest_offloads(struct virtnet_info *vi) 3328 { 3329 u64 offloads = vi->guest_offloads; 3330 3331 if (!vi->guest_offloads) 3332 return 0; 3333 3334 return virtnet_set_guest_offloads(vi, offloads); 3335 } 3336 3337 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog, 3338 struct netlink_ext_ack *extack) 3339 { 3340 unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM + 3341 sizeof(struct skb_shared_info)); 3342 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN; 3343 struct virtnet_info *vi = netdev_priv(dev); 3344 struct bpf_prog *old_prog; 3345 u16 xdp_qp = 0, curr_qp; 3346 int i, err; 3347 3348 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) 3349 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 3350 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 3351 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 3352 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) || 3353 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) || 3354 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) || 3355 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) { 3356 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first"); 3357 return -EOPNOTSUPP; 3358 } 3359 3360 if (vi->mergeable_rx_bufs && !vi->any_header_sg) { 3361 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required"); 3362 return -EINVAL; 3363 } 3364 3365 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) { 3366 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags"); 3367 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz); 3368 return -EINVAL; 3369 } 3370 3371 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs; 3372 if (prog) 3373 xdp_qp = nr_cpu_ids; 3374 3375 /* XDP requires extra queues for XDP_TX */ 3376 if (curr_qp + xdp_qp > vi->max_queue_pairs) { 3377 netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n", 3378 curr_qp + xdp_qp, vi->max_queue_pairs); 3379 xdp_qp = 0; 3380 } 3381 3382 old_prog = rtnl_dereference(vi->rq[0].xdp_prog); 3383 if (!prog && !old_prog) 3384 return 0; 3385 3386 if (prog) 3387 bpf_prog_add(prog, vi->max_queue_pairs - 1); 3388 3389 /* Make sure NAPI is not using any XDP TX queues for RX. */ 3390 if (netif_running(dev)) { 3391 for (i = 0; i < vi->max_queue_pairs; i++) { 3392 napi_disable(&vi->rq[i].napi); 3393 virtnet_napi_tx_disable(&vi->sq[i].napi); 3394 } 3395 } 3396 3397 if (!prog) { 3398 for (i = 0; i < vi->max_queue_pairs; i++) { 3399 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 3400 if (i == 0) 3401 virtnet_restore_guest_offloads(vi); 3402 } 3403 synchronize_net(); 3404 } 3405 3406 err = _virtnet_set_queues(vi, curr_qp + xdp_qp); 3407 if (err) 3408 goto err; 3409 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp); 3410 vi->xdp_queue_pairs = xdp_qp; 3411 3412 if (prog) { 3413 vi->xdp_enabled = true; 3414 for (i = 0; i < vi->max_queue_pairs; i++) { 3415 rcu_assign_pointer(vi->rq[i].xdp_prog, prog); 3416 if (i == 0 && !old_prog) 3417 virtnet_clear_guest_offloads(vi); 3418 } 3419 if (!old_prog) 3420 xdp_features_set_redirect_target(dev, true); 3421 } else { 3422 xdp_features_clear_redirect_target(dev); 3423 vi->xdp_enabled = false; 3424 } 3425 3426 for (i = 0; i < vi->max_queue_pairs; i++) { 3427 if (old_prog) 3428 bpf_prog_put(old_prog); 3429 if (netif_running(dev)) { 3430 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 3431 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 3432 &vi->sq[i].napi); 3433 } 3434 } 3435 3436 return 0; 3437 3438 err: 3439 if (!prog) { 3440 virtnet_clear_guest_offloads(vi); 3441 for (i = 0; i < vi->max_queue_pairs; i++) 3442 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog); 3443 } 3444 3445 if (netif_running(dev)) { 3446 for (i = 0; i < vi->max_queue_pairs; i++) { 3447 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi); 3448 virtnet_napi_tx_enable(vi, vi->sq[i].vq, 3449 &vi->sq[i].napi); 3450 } 3451 } 3452 if (prog) 3453 bpf_prog_sub(prog, vi->max_queue_pairs - 1); 3454 return err; 3455 } 3456 3457 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp) 3458 { 3459 switch (xdp->command) { 3460 case XDP_SETUP_PROG: 3461 return virtnet_xdp_set(dev, xdp->prog, xdp->extack); 3462 default: 3463 return -EINVAL; 3464 } 3465 } 3466 3467 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf, 3468 size_t len) 3469 { 3470 struct virtnet_info *vi = netdev_priv(dev); 3471 int ret; 3472 3473 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY)) 3474 return -EOPNOTSUPP; 3475 3476 ret = snprintf(buf, len, "sby"); 3477 if (ret >= len) 3478 return -EOPNOTSUPP; 3479 3480 return 0; 3481 } 3482 3483 static int virtnet_set_features(struct net_device *dev, 3484 netdev_features_t features) 3485 { 3486 struct virtnet_info *vi = netdev_priv(dev); 3487 u64 offloads; 3488 int err; 3489 3490 if ((dev->features ^ features) & NETIF_F_GRO_HW) { 3491 if (vi->xdp_enabled) 3492 return -EBUSY; 3493 3494 if (features & NETIF_F_GRO_HW) 3495 offloads = vi->guest_offloads_capable; 3496 else 3497 offloads = vi->guest_offloads_capable & 3498 ~GUEST_OFFLOAD_GRO_HW_MASK; 3499 3500 err = virtnet_set_guest_offloads(vi, offloads); 3501 if (err) 3502 return err; 3503 vi->guest_offloads = offloads; 3504 } 3505 3506 if ((dev->features ^ features) & NETIF_F_RXHASH) { 3507 if (features & NETIF_F_RXHASH) 3508 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved; 3509 else 3510 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE; 3511 3512 if (!virtnet_commit_rss_command(vi)) 3513 return -EINVAL; 3514 } 3515 3516 return 0; 3517 } 3518 3519 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue) 3520 { 3521 struct virtnet_info *priv = netdev_priv(dev); 3522 struct send_queue *sq = &priv->sq[txqueue]; 3523 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue); 3524 3525 u64_stats_update_begin(&sq->stats.syncp); 3526 sq->stats.tx_timeouts++; 3527 u64_stats_update_end(&sq->stats.syncp); 3528 3529 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n", 3530 txqueue, sq->name, sq->vq->index, sq->vq->name, 3531 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start))); 3532 } 3533 3534 static const struct net_device_ops virtnet_netdev = { 3535 .ndo_open = virtnet_open, 3536 .ndo_stop = virtnet_close, 3537 .ndo_start_xmit = start_xmit, 3538 .ndo_validate_addr = eth_validate_addr, 3539 .ndo_set_mac_address = virtnet_set_mac_address, 3540 .ndo_set_rx_mode = virtnet_set_rx_mode, 3541 .ndo_get_stats64 = virtnet_stats, 3542 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid, 3543 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid, 3544 .ndo_bpf = virtnet_xdp, 3545 .ndo_xdp_xmit = virtnet_xdp_xmit, 3546 .ndo_features_check = passthru_features_check, 3547 .ndo_get_phys_port_name = virtnet_get_phys_port_name, 3548 .ndo_set_features = virtnet_set_features, 3549 .ndo_tx_timeout = virtnet_tx_timeout, 3550 }; 3551 3552 static void virtnet_config_changed_work(struct work_struct *work) 3553 { 3554 struct virtnet_info *vi = 3555 container_of(work, struct virtnet_info, config_work); 3556 u16 v; 3557 3558 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS, 3559 struct virtio_net_config, status, &v) < 0) 3560 return; 3561 3562 if (v & VIRTIO_NET_S_ANNOUNCE) { 3563 netdev_notify_peers(vi->dev); 3564 virtnet_ack_link_announce(vi); 3565 } 3566 3567 /* Ignore unknown (future) status bits */ 3568 v &= VIRTIO_NET_S_LINK_UP; 3569 3570 if (vi->status == v) 3571 return; 3572 3573 vi->status = v; 3574 3575 if (vi->status & VIRTIO_NET_S_LINK_UP) { 3576 virtnet_update_settings(vi); 3577 netif_carrier_on(vi->dev); 3578 netif_tx_wake_all_queues(vi->dev); 3579 } else { 3580 netif_carrier_off(vi->dev); 3581 netif_tx_stop_all_queues(vi->dev); 3582 } 3583 } 3584 3585 static void virtnet_config_changed(struct virtio_device *vdev) 3586 { 3587 struct virtnet_info *vi = vdev->priv; 3588 3589 schedule_work(&vi->config_work); 3590 } 3591 3592 static void virtnet_free_queues(struct virtnet_info *vi) 3593 { 3594 int i; 3595 3596 for (i = 0; i < vi->max_queue_pairs; i++) { 3597 __netif_napi_del(&vi->rq[i].napi); 3598 __netif_napi_del(&vi->sq[i].napi); 3599 } 3600 3601 /* We called __netif_napi_del(), 3602 * we need to respect an RCU grace period before freeing vi->rq 3603 */ 3604 synchronize_net(); 3605 3606 kfree(vi->rq); 3607 kfree(vi->sq); 3608 kfree(vi->ctrl); 3609 } 3610 3611 static void _free_receive_bufs(struct virtnet_info *vi) 3612 { 3613 struct bpf_prog *old_prog; 3614 int i; 3615 3616 for (i = 0; i < vi->max_queue_pairs; i++) { 3617 while (vi->rq[i].pages) 3618 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0); 3619 3620 old_prog = rtnl_dereference(vi->rq[i].xdp_prog); 3621 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL); 3622 if (old_prog) 3623 bpf_prog_put(old_prog); 3624 } 3625 } 3626 3627 static void free_receive_bufs(struct virtnet_info *vi) 3628 { 3629 rtnl_lock(); 3630 _free_receive_bufs(vi); 3631 rtnl_unlock(); 3632 } 3633 3634 static void free_receive_page_frags(struct virtnet_info *vi) 3635 { 3636 int i; 3637 for (i = 0; i < vi->max_queue_pairs; i++) 3638 if (vi->rq[i].alloc_frag.page) 3639 put_page(vi->rq[i].alloc_frag.page); 3640 } 3641 3642 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf) 3643 { 3644 if (!is_xdp_frame(buf)) 3645 dev_kfree_skb(buf); 3646 else 3647 xdp_return_frame(ptr_to_xdp(buf)); 3648 } 3649 3650 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf) 3651 { 3652 struct virtnet_info *vi = vq->vdev->priv; 3653 int i = vq2rxq(vq); 3654 3655 if (vi->mergeable_rx_bufs) 3656 put_page(virt_to_head_page(buf)); 3657 else if (vi->big_packets) 3658 give_pages(&vi->rq[i], buf); 3659 else 3660 put_page(virt_to_head_page(buf)); 3661 } 3662 3663 static void free_unused_bufs(struct virtnet_info *vi) 3664 { 3665 void *buf; 3666 int i; 3667 3668 for (i = 0; i < vi->max_queue_pairs; i++) { 3669 struct virtqueue *vq = vi->sq[i].vq; 3670 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) 3671 virtnet_sq_free_unused_buf(vq, buf); 3672 cond_resched(); 3673 } 3674 3675 for (i = 0; i < vi->max_queue_pairs; i++) { 3676 struct virtqueue *vq = vi->rq[i].vq; 3677 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) 3678 virtnet_rq_free_unused_buf(vq, buf); 3679 cond_resched(); 3680 } 3681 } 3682 3683 static void virtnet_del_vqs(struct virtnet_info *vi) 3684 { 3685 struct virtio_device *vdev = vi->vdev; 3686 3687 virtnet_clean_affinity(vi); 3688 3689 vdev->config->del_vqs(vdev); 3690 3691 virtnet_free_queues(vi); 3692 } 3693 3694 /* How large should a single buffer be so a queue full of these can fit at 3695 * least one full packet? 3696 * Logic below assumes the mergeable buffer header is used. 3697 */ 3698 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq) 3699 { 3700 const unsigned int hdr_len = vi->hdr_len; 3701 unsigned int rq_size = virtqueue_get_vring_size(vq); 3702 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu; 3703 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len; 3704 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size); 3705 3706 return max(max(min_buf_len, hdr_len) - hdr_len, 3707 (unsigned int)GOOD_PACKET_LEN); 3708 } 3709 3710 static int virtnet_find_vqs(struct virtnet_info *vi) 3711 { 3712 vq_callback_t **callbacks; 3713 struct virtqueue **vqs; 3714 int ret = -ENOMEM; 3715 int i, total_vqs; 3716 const char **names; 3717 bool *ctx; 3718 3719 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by 3720 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by 3721 * possible control vq. 3722 */ 3723 total_vqs = vi->max_queue_pairs * 2 + 3724 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ); 3725 3726 /* Allocate space for find_vqs parameters */ 3727 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL); 3728 if (!vqs) 3729 goto err_vq; 3730 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL); 3731 if (!callbacks) 3732 goto err_callback; 3733 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL); 3734 if (!names) 3735 goto err_names; 3736 if (!vi->big_packets || vi->mergeable_rx_bufs) { 3737 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL); 3738 if (!ctx) 3739 goto err_ctx; 3740 } else { 3741 ctx = NULL; 3742 } 3743 3744 /* Parameters for control virtqueue, if any */ 3745 if (vi->has_cvq) { 3746 callbacks[total_vqs - 1] = NULL; 3747 names[total_vqs - 1] = "control"; 3748 } 3749 3750 /* Allocate/initialize parameters for send/receive virtqueues */ 3751 for (i = 0; i < vi->max_queue_pairs; i++) { 3752 callbacks[rxq2vq(i)] = skb_recv_done; 3753 callbacks[txq2vq(i)] = skb_xmit_done; 3754 sprintf(vi->rq[i].name, "input.%d", i); 3755 sprintf(vi->sq[i].name, "output.%d", i); 3756 names[rxq2vq(i)] = vi->rq[i].name; 3757 names[txq2vq(i)] = vi->sq[i].name; 3758 if (ctx) 3759 ctx[rxq2vq(i)] = true; 3760 } 3761 3762 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks, 3763 names, ctx, NULL); 3764 if (ret) 3765 goto err_find; 3766 3767 if (vi->has_cvq) { 3768 vi->cvq = vqs[total_vqs - 1]; 3769 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN)) 3770 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER; 3771 } 3772 3773 for (i = 0; i < vi->max_queue_pairs; i++) { 3774 vi->rq[i].vq = vqs[rxq2vq(i)]; 3775 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq); 3776 vi->sq[i].vq = vqs[txq2vq(i)]; 3777 } 3778 3779 /* run here: ret == 0. */ 3780 3781 3782 err_find: 3783 kfree(ctx); 3784 err_ctx: 3785 kfree(names); 3786 err_names: 3787 kfree(callbacks); 3788 err_callback: 3789 kfree(vqs); 3790 err_vq: 3791 return ret; 3792 } 3793 3794 static int virtnet_alloc_queues(struct virtnet_info *vi) 3795 { 3796 int i; 3797 3798 if (vi->has_cvq) { 3799 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL); 3800 if (!vi->ctrl) 3801 goto err_ctrl; 3802 } else { 3803 vi->ctrl = NULL; 3804 } 3805 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL); 3806 if (!vi->sq) 3807 goto err_sq; 3808 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL); 3809 if (!vi->rq) 3810 goto err_rq; 3811 3812 INIT_DELAYED_WORK(&vi->refill, refill_work); 3813 for (i = 0; i < vi->max_queue_pairs; i++) { 3814 vi->rq[i].pages = NULL; 3815 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll, 3816 napi_weight); 3817 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi, 3818 virtnet_poll_tx, 3819 napi_tx ? napi_weight : 0); 3820 3821 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg)); 3822 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len); 3823 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg)); 3824 3825 u64_stats_init(&vi->rq[i].stats.syncp); 3826 u64_stats_init(&vi->sq[i].stats.syncp); 3827 } 3828 3829 return 0; 3830 3831 err_rq: 3832 kfree(vi->sq); 3833 err_sq: 3834 kfree(vi->ctrl); 3835 err_ctrl: 3836 return -ENOMEM; 3837 } 3838 3839 static int init_vqs(struct virtnet_info *vi) 3840 { 3841 int ret; 3842 3843 /* Allocate send & receive queues */ 3844 ret = virtnet_alloc_queues(vi); 3845 if (ret) 3846 goto err; 3847 3848 ret = virtnet_find_vqs(vi); 3849 if (ret) 3850 goto err_free; 3851 3852 cpus_read_lock(); 3853 virtnet_set_affinity(vi); 3854 cpus_read_unlock(); 3855 3856 return 0; 3857 3858 err_free: 3859 virtnet_free_queues(vi); 3860 err: 3861 return ret; 3862 } 3863 3864 #ifdef CONFIG_SYSFS 3865 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue, 3866 char *buf) 3867 { 3868 struct virtnet_info *vi = netdev_priv(queue->dev); 3869 unsigned int queue_index = get_netdev_rx_queue_index(queue); 3870 unsigned int headroom = virtnet_get_headroom(vi); 3871 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0; 3872 struct ewma_pkt_len *avg; 3873 3874 BUG_ON(queue_index >= vi->max_queue_pairs); 3875 avg = &vi->rq[queue_index].mrg_avg_pkt_len; 3876 return sprintf(buf, "%u\n", 3877 get_mergeable_buf_len(&vi->rq[queue_index], avg, 3878 SKB_DATA_ALIGN(headroom + tailroom))); 3879 } 3880 3881 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute = 3882 __ATTR_RO(mergeable_rx_buffer_size); 3883 3884 static struct attribute *virtio_net_mrg_rx_attrs[] = { 3885 &mergeable_rx_buffer_size_attribute.attr, 3886 NULL 3887 }; 3888 3889 static const struct attribute_group virtio_net_mrg_rx_group = { 3890 .name = "virtio_net", 3891 .attrs = virtio_net_mrg_rx_attrs 3892 }; 3893 #endif 3894 3895 static bool virtnet_fail_on_feature(struct virtio_device *vdev, 3896 unsigned int fbit, 3897 const char *fname, const char *dname) 3898 { 3899 if (!virtio_has_feature(vdev, fbit)) 3900 return false; 3901 3902 dev_err(&vdev->dev, "device advertises feature %s but not %s", 3903 fname, dname); 3904 3905 return true; 3906 } 3907 3908 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \ 3909 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit) 3910 3911 static bool virtnet_validate_features(struct virtio_device *vdev) 3912 { 3913 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) && 3914 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX, 3915 "VIRTIO_NET_F_CTRL_VQ") || 3916 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN, 3917 "VIRTIO_NET_F_CTRL_VQ") || 3918 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE, 3919 "VIRTIO_NET_F_CTRL_VQ") || 3920 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") || 3921 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR, 3922 "VIRTIO_NET_F_CTRL_VQ") || 3923 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS, 3924 "VIRTIO_NET_F_CTRL_VQ") || 3925 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT, 3926 "VIRTIO_NET_F_CTRL_VQ") || 3927 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL, 3928 "VIRTIO_NET_F_CTRL_VQ"))) { 3929 return false; 3930 } 3931 3932 return true; 3933 } 3934 3935 #define MIN_MTU ETH_MIN_MTU 3936 #define MAX_MTU ETH_MAX_MTU 3937 3938 static int virtnet_validate(struct virtio_device *vdev) 3939 { 3940 if (!vdev->config->get) { 3941 dev_err(&vdev->dev, "%s failure: config access disabled\n", 3942 __func__); 3943 return -EINVAL; 3944 } 3945 3946 if (!virtnet_validate_features(vdev)) 3947 return -EINVAL; 3948 3949 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 3950 int mtu = virtio_cread16(vdev, 3951 offsetof(struct virtio_net_config, 3952 mtu)); 3953 if (mtu < MIN_MTU) 3954 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU); 3955 } 3956 3957 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) && 3958 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) { 3959 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby"); 3960 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY); 3961 } 3962 3963 return 0; 3964 } 3965 3966 static bool virtnet_check_guest_gso(const struct virtnet_info *vi) 3967 { 3968 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) || 3969 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) || 3970 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) || 3971 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) || 3972 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) && 3973 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6)); 3974 } 3975 3976 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu) 3977 { 3978 bool guest_gso = virtnet_check_guest_gso(vi); 3979 3980 /* If device can receive ANY guest GSO packets, regardless of mtu, 3981 * allocate packets of maximum size, otherwise limit it to only 3982 * mtu size worth only. 3983 */ 3984 if (mtu > ETH_DATA_LEN || guest_gso) { 3985 vi->big_packets = true; 3986 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE); 3987 } 3988 } 3989 3990 static int virtnet_probe(struct virtio_device *vdev) 3991 { 3992 int i, err = -ENOMEM; 3993 struct net_device *dev; 3994 struct virtnet_info *vi; 3995 u16 max_queue_pairs; 3996 int mtu = 0; 3997 3998 /* Find if host supports multiqueue/rss virtio_net device */ 3999 max_queue_pairs = 1; 4000 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 4001 max_queue_pairs = 4002 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs)); 4003 4004 /* We need at least 2 queue's */ 4005 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN || 4006 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX || 4007 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 4008 max_queue_pairs = 1; 4009 4010 /* Allocate ourselves a network device with room for our info */ 4011 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs); 4012 if (!dev) 4013 return -ENOMEM; 4014 4015 /* Set up network device as normal. */ 4016 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE | 4017 IFF_TX_SKB_NO_LINEAR; 4018 dev->netdev_ops = &virtnet_netdev; 4019 dev->features = NETIF_F_HIGHDMA; 4020 4021 dev->ethtool_ops = &virtnet_ethtool_ops; 4022 SET_NETDEV_DEV(dev, &vdev->dev); 4023 4024 /* Do we support "hardware" checksums? */ 4025 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) { 4026 /* This opens up the world of extra features. */ 4027 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG; 4028 if (csum) 4029 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG; 4030 4031 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) { 4032 dev->hw_features |= NETIF_F_TSO 4033 | NETIF_F_TSO_ECN | NETIF_F_TSO6; 4034 } 4035 /* Individual feature bits: what can host handle? */ 4036 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4)) 4037 dev->hw_features |= NETIF_F_TSO; 4038 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6)) 4039 dev->hw_features |= NETIF_F_TSO6; 4040 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN)) 4041 dev->hw_features |= NETIF_F_TSO_ECN; 4042 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO)) 4043 dev->hw_features |= NETIF_F_GSO_UDP_L4; 4044 4045 dev->features |= NETIF_F_GSO_ROBUST; 4046 4047 if (gso) 4048 dev->features |= dev->hw_features & NETIF_F_ALL_TSO; 4049 /* (!csum && gso) case will be fixed by register_netdev() */ 4050 } 4051 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM)) 4052 dev->features |= NETIF_F_RXCSUM; 4053 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) || 4054 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6)) 4055 dev->features |= NETIF_F_GRO_HW; 4056 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) 4057 dev->hw_features |= NETIF_F_GRO_HW; 4058 4059 dev->vlan_features = dev->features; 4060 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT; 4061 4062 /* MTU range: 68 - 65535 */ 4063 dev->min_mtu = MIN_MTU; 4064 dev->max_mtu = MAX_MTU; 4065 4066 /* Configuration may specify what MAC to use. Otherwise random. */ 4067 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) { 4068 u8 addr[ETH_ALEN]; 4069 4070 virtio_cread_bytes(vdev, 4071 offsetof(struct virtio_net_config, mac), 4072 addr, ETH_ALEN); 4073 eth_hw_addr_set(dev, addr); 4074 } else { 4075 eth_hw_addr_random(dev); 4076 dev_info(&vdev->dev, "Assigned random MAC address %pM\n", 4077 dev->dev_addr); 4078 } 4079 4080 /* Set up our device-specific information */ 4081 vi = netdev_priv(dev); 4082 vi->dev = dev; 4083 vi->vdev = vdev; 4084 vdev->priv = vi; 4085 4086 INIT_WORK(&vi->config_work, virtnet_config_changed_work); 4087 spin_lock_init(&vi->refill_lock); 4088 4089 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) { 4090 vi->mergeable_rx_bufs = true; 4091 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG; 4092 } 4093 4094 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) { 4095 vi->rx_usecs = 0; 4096 vi->tx_usecs = 0; 4097 vi->tx_max_packets = 0; 4098 vi->rx_max_packets = 0; 4099 } 4100 4101 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT)) 4102 vi->has_rss_hash_report = true; 4103 4104 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS)) 4105 vi->has_rss = true; 4106 4107 if (vi->has_rss || vi->has_rss_hash_report) { 4108 vi->rss_indir_table_size = 4109 virtio_cread16(vdev, offsetof(struct virtio_net_config, 4110 rss_max_indirection_table_length)); 4111 vi->rss_key_size = 4112 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size)); 4113 4114 vi->rss_hash_types_supported = 4115 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types)); 4116 vi->rss_hash_types_supported &= 4117 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX | 4118 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX | 4119 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX); 4120 4121 dev->hw_features |= NETIF_F_RXHASH; 4122 } 4123 4124 if (vi->has_rss_hash_report) 4125 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash); 4126 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) || 4127 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 4128 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf); 4129 else 4130 vi->hdr_len = sizeof(struct virtio_net_hdr); 4131 4132 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) || 4133 virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) 4134 vi->any_header_sg = true; 4135 4136 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ)) 4137 vi->has_cvq = true; 4138 4139 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) { 4140 mtu = virtio_cread16(vdev, 4141 offsetof(struct virtio_net_config, 4142 mtu)); 4143 if (mtu < dev->min_mtu) { 4144 /* Should never trigger: MTU was previously validated 4145 * in virtnet_validate. 4146 */ 4147 dev_err(&vdev->dev, 4148 "device MTU appears to have changed it is now %d < %d", 4149 mtu, dev->min_mtu); 4150 err = -EINVAL; 4151 goto free; 4152 } 4153 4154 dev->mtu = mtu; 4155 dev->max_mtu = mtu; 4156 } 4157 4158 virtnet_set_big_packets(vi, mtu); 4159 4160 if (vi->any_header_sg) 4161 dev->needed_headroom = vi->hdr_len; 4162 4163 /* Enable multiqueue by default */ 4164 if (num_online_cpus() >= max_queue_pairs) 4165 vi->curr_queue_pairs = max_queue_pairs; 4166 else 4167 vi->curr_queue_pairs = num_online_cpus(); 4168 vi->max_queue_pairs = max_queue_pairs; 4169 4170 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */ 4171 err = init_vqs(vi); 4172 if (err) 4173 goto free; 4174 4175 #ifdef CONFIG_SYSFS 4176 if (vi->mergeable_rx_bufs) 4177 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group; 4178 #endif 4179 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs); 4180 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs); 4181 4182 virtnet_init_settings(dev); 4183 4184 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) { 4185 vi->failover = net_failover_create(vi->dev); 4186 if (IS_ERR(vi->failover)) { 4187 err = PTR_ERR(vi->failover); 4188 goto free_vqs; 4189 } 4190 } 4191 4192 if (vi->has_rss || vi->has_rss_hash_report) 4193 virtnet_init_default_rss(vi); 4194 4195 /* serialize netdev register + virtio_device_ready() with ndo_open() */ 4196 rtnl_lock(); 4197 4198 err = register_netdevice(dev); 4199 if (err) { 4200 pr_debug("virtio_net: registering device failed\n"); 4201 rtnl_unlock(); 4202 goto free_failover; 4203 } 4204 4205 virtio_device_ready(vdev); 4206 4207 /* a random MAC address has been assigned, notify the device. 4208 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there 4209 * because many devices work fine without getting MAC explicitly 4210 */ 4211 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) && 4212 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 4213 struct scatterlist sg; 4214 4215 sg_init_one(&sg, dev->dev_addr, dev->addr_len); 4216 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC, 4217 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) { 4218 pr_debug("virtio_net: setting MAC address failed\n"); 4219 rtnl_unlock(); 4220 err = -EINVAL; 4221 goto free_unregister_netdev; 4222 } 4223 } 4224 4225 rtnl_unlock(); 4226 4227 err = virtnet_cpu_notif_add(vi); 4228 if (err) { 4229 pr_debug("virtio_net: registering cpu notifier failed\n"); 4230 goto free_unregister_netdev; 4231 } 4232 4233 virtnet_set_queues(vi, vi->curr_queue_pairs); 4234 4235 /* Assume link up if device can't report link status, 4236 otherwise get link status from config. */ 4237 netif_carrier_off(dev); 4238 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) { 4239 schedule_work(&vi->config_work); 4240 } else { 4241 vi->status = VIRTIO_NET_S_LINK_UP; 4242 virtnet_update_settings(vi); 4243 netif_carrier_on(dev); 4244 } 4245 4246 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++) 4247 if (virtio_has_feature(vi->vdev, guest_offloads[i])) 4248 set_bit(guest_offloads[i], &vi->guest_offloads); 4249 vi->guest_offloads_capable = vi->guest_offloads; 4250 4251 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n", 4252 dev->name, max_queue_pairs); 4253 4254 return 0; 4255 4256 free_unregister_netdev: 4257 unregister_netdev(dev); 4258 free_failover: 4259 net_failover_destroy(vi->failover); 4260 free_vqs: 4261 virtio_reset_device(vdev); 4262 cancel_delayed_work_sync(&vi->refill); 4263 free_receive_page_frags(vi); 4264 virtnet_del_vqs(vi); 4265 free: 4266 free_netdev(dev); 4267 return err; 4268 } 4269 4270 static void remove_vq_common(struct virtnet_info *vi) 4271 { 4272 virtio_reset_device(vi->vdev); 4273 4274 /* Free unused buffers in both send and recv, if any. */ 4275 free_unused_bufs(vi); 4276 4277 free_receive_bufs(vi); 4278 4279 free_receive_page_frags(vi); 4280 4281 virtnet_del_vqs(vi); 4282 } 4283 4284 static void virtnet_remove(struct virtio_device *vdev) 4285 { 4286 struct virtnet_info *vi = vdev->priv; 4287 4288 virtnet_cpu_notif_remove(vi); 4289 4290 /* Make sure no work handler is accessing the device. */ 4291 flush_work(&vi->config_work); 4292 4293 unregister_netdev(vi->dev); 4294 4295 net_failover_destroy(vi->failover); 4296 4297 remove_vq_common(vi); 4298 4299 free_netdev(vi->dev); 4300 } 4301 4302 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev) 4303 { 4304 struct virtnet_info *vi = vdev->priv; 4305 4306 virtnet_cpu_notif_remove(vi); 4307 virtnet_freeze_down(vdev); 4308 remove_vq_common(vi); 4309 4310 return 0; 4311 } 4312 4313 static __maybe_unused int virtnet_restore(struct virtio_device *vdev) 4314 { 4315 struct virtnet_info *vi = vdev->priv; 4316 int err; 4317 4318 err = virtnet_restore_up(vdev); 4319 if (err) 4320 return err; 4321 virtnet_set_queues(vi, vi->curr_queue_pairs); 4322 4323 err = virtnet_cpu_notif_add(vi); 4324 if (err) { 4325 virtnet_freeze_down(vdev); 4326 remove_vq_common(vi); 4327 return err; 4328 } 4329 4330 return 0; 4331 } 4332 4333 static struct virtio_device_id id_table[] = { 4334 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID }, 4335 { 0 }, 4336 }; 4337 4338 #define VIRTNET_FEATURES \ 4339 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \ 4340 VIRTIO_NET_F_MAC, \ 4341 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \ 4342 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \ 4343 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \ 4344 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \ 4345 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \ 4346 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \ 4347 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \ 4348 VIRTIO_NET_F_CTRL_MAC_ADDR, \ 4349 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \ 4350 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \ 4351 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \ 4352 VIRTIO_NET_F_GUEST_HDRLEN 4353 4354 static unsigned int features[] = { 4355 VIRTNET_FEATURES, 4356 }; 4357 4358 static unsigned int features_legacy[] = { 4359 VIRTNET_FEATURES, 4360 VIRTIO_NET_F_GSO, 4361 VIRTIO_F_ANY_LAYOUT, 4362 }; 4363 4364 static struct virtio_driver virtio_net_driver = { 4365 .feature_table = features, 4366 .feature_table_size = ARRAY_SIZE(features), 4367 .feature_table_legacy = features_legacy, 4368 .feature_table_size_legacy = ARRAY_SIZE(features_legacy), 4369 .driver.name = KBUILD_MODNAME, 4370 .driver.owner = THIS_MODULE, 4371 .id_table = id_table, 4372 .validate = virtnet_validate, 4373 .probe = virtnet_probe, 4374 .remove = virtnet_remove, 4375 .config_changed = virtnet_config_changed, 4376 #ifdef CONFIG_PM_SLEEP 4377 .freeze = virtnet_freeze, 4378 .restore = virtnet_restore, 4379 #endif 4380 }; 4381 4382 static __init int virtio_net_driver_init(void) 4383 { 4384 int ret; 4385 4386 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online", 4387 virtnet_cpu_online, 4388 virtnet_cpu_down_prep); 4389 if (ret < 0) 4390 goto out; 4391 virtionet_online = ret; 4392 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead", 4393 NULL, virtnet_cpu_dead); 4394 if (ret) 4395 goto err_dead; 4396 ret = register_virtio_driver(&virtio_net_driver); 4397 if (ret) 4398 goto err_virtio; 4399 return 0; 4400 err_virtio: 4401 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 4402 err_dead: 4403 cpuhp_remove_multi_state(virtionet_online); 4404 out: 4405 return ret; 4406 } 4407 module_init(virtio_net_driver_init); 4408 4409 static __exit void virtio_net_driver_exit(void) 4410 { 4411 unregister_virtio_driver(&virtio_net_driver); 4412 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD); 4413 cpuhp_remove_multi_state(virtionet_online); 4414 } 4415 module_exit(virtio_net_driver_exit); 4416 4417 MODULE_DEVICE_TABLE(virtio, id_table); 4418 MODULE_DESCRIPTION("Virtio network driver"); 4419 MODULE_LICENSE("GPL"); 4420